Merge branch 'master' into master

modules/api/api.py

@@ -18,7 +18,8 @@ from modules.textual_inversion.textual_inversion import create_embedding, train_
 from modules.textual_inversion.preprocess import preprocess
 from modules.hypernetworks.hypernetwork import create_hypernetwork, train_hypernetwork
 from PIL import PngImagePlugin,Image
-from modules.sd_models import checkpoints_list, find_checkpoint_config
+from modules.sd_models import checkpoints_list
+from modules.sd_models_config import find_checkpoint_config_near_filename
 from modules.realesrgan_model import get_realesrgan_models
 from modules import devices
 from typing import List
@@ -387,7 +388,7 @@ class Api:
         ]
 
     def get_sd_models(self):
-        return [{"title": x.title, "model_name": x.model_name, "hash": x.shorthash, "sha256": x.sha256, "filename": x.filename, "config": find_checkpoint_config(x)} for x in checkpoints_list.values()]
+        return [{"title": x.title, "model_name": x.model_name, "hash": x.shorthash, "sha256": x.sha256, "filename": x.filename, "config": find_checkpoint_config_near_filename(x)} for x in checkpoints_list.values()]
 
     def get_hypernetworks(self):
         return [{"name": name, "path": shared.hypernetworks[name]} for name in shared.hypernetworks]

modules/api/models.py

@@ -228,7 +228,7 @@ class SDModelItem(BaseModel):
     hash: Optional[str] = Field(title="Short hash")
     sha256: Optional[str] = Field(title="sha256 hash")
     filename: str = Field(title="Filename")
-    config: str = Field(title="Config file")
+    config: Optional[str] = Field(title="Config file")
 
 class HypernetworkItem(BaseModel):
     name: str = Field(title="Name")

modules/codeformer_model.py

@@ -8,7 +8,7 @@ import torch
 import modules.face_restoration
 import modules.shared
 from modules import shared, devices, modelloader
-from modules.paths import script_path, models_path
+from modules.paths import models_path
 
 # codeformer people made a choice to include modified basicsr library to their project which makes
 # it utterly impossible to use it alongside with other libraries that also use basicsr, like GFPGAN.

modules/deepbooru_model.py

@@ -2,6 +2,8 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+from modules import devices
+
 # see https://github.com/AUTOMATIC1111/TorchDeepDanbooru for more
 
 
@@ -196,7 +198,7 @@ class DeepDanbooruModel(nn.Module):
         t_358, = inputs
         t_359 = t_358.permute(*[0, 3, 1, 2])
         t_359_padded = F.pad(t_359, [2, 3, 2, 3], value=0)
-        t_360 = self.n_Conv_0(t_359_padded)
+        t_360 = self.n_Conv_0(t_359_padded.to(self.n_Conv_0.bias.dtype) if devices.unet_needs_upcast else t_359_padded)
         t_361 = F.relu(t_360)
         t_361 = F.pad(t_361, [0, 1, 0, 1], value=float('-inf'))
         t_362 = self.n_MaxPool_0(t_361)

modules/devices.py

@@ -34,14 +34,18 @@ def get_cuda_device_string():
     return "cuda"
 
 
-def get_optimal_device():
+def get_optimal_device_name():
     if torch.cuda.is_available():
-        return torch.device(get_cuda_device_string())
+        return get_cuda_device_string()
 
     if has_mps():
-        return torch.device("mps")
+        return "mps"
 
-    return cpu
+    return "cpu"
+
+
+def get_optimal_device():
+    return torch.device(get_optimal_device_name())
 
 
 def get_device_for(task):
@@ -79,6 +83,16 @@ cpu = torch.device("cpu")
 device = device_interrogate = device_gfpgan = device_esrgan = device_codeformer = None
 dtype = torch.float16
 dtype_vae = torch.float16
+dtype_unet = torch.float16
+unet_needs_upcast = False
+
+
+def cond_cast_unet(input):
+    return input.to(dtype_unet) if unet_needs_upcast else input
+
+
+def cond_cast_float(input):
+    return input.float() if unet_needs_upcast else input
 
 
 def randn(seed, shape):
@@ -106,6 +120,10 @@ def autocast(disable=False):
     return torch.autocast("cuda")
 
 
+def without_autocast(disable=False):
+    return torch.autocast("cuda", enabled=False) if torch.is_autocast_enabled() and not disable else contextlib.nullcontext()
+
+
 class NansException(Exception):
     pass
 
@@ -123,7 +141,7 @@ def test_for_nans(x, where):
         message = "A tensor with all NaNs was produced in Unet."
 
         if not shared.cmd_opts.no_half:
-            message += " This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try using --no-half commandline argument to fix this."
+            message += " This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the \"Upcast cross attention layer to float32\" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this."
 
     elif where == "vae":
         message = "A tensor with all NaNs was produced in VAE."
@@ -133,6 +151,8 @@ def test_for_nans(x, where):
     else:
         message = "A tensor with all NaNs was produced."
 
+    message += " Use --disable-nan-check commandline argument to disable this check."
+
     raise NansException(message)
 
 
@@ -187,6 +207,3 @@ if has_mps():
         cumsum_needs_bool_fix = not torch.BoolTensor([True,True]).to(device=torch.device("mps"), dtype=torch.int64).equal(torch.BoolTensor([True,False]).to(torch.device("mps")).cumsum(0))
         torch.cumsum = lambda input, *args, **kwargs: ( cumsum_fix(input, orig_cumsum, *args, **kwargs) )
         torch.Tensor.cumsum = lambda self, *args, **kwargs: ( cumsum_fix(self, orig_Tensor_cumsum, *args, **kwargs) )
-        orig_narrow = torch.narrow
-        torch.narrow = lambda *args, **kwargs: ( orig_narrow(*args, **kwargs).clone() )
-

modules/extensions.py

@@ -7,9 +7,11 @@ import git
 from modules import paths, shared
 
 extensions = []
-extensions_dir = os.path.join(paths.script_path, "extensions")
+extensions_dir = os.path.join(paths.data_path, "extensions")
 extensions_builtin_dir = os.path.join(paths.script_path, "extensions-builtin")
 
+if not os.path.exists(extensions_dir):
+    os.makedirs(extensions_dir)
 
 def active():
     return [x for x in extensions if x.enabled]

modules/extra_networks_hypernet.py

@@ -1,4 +1,4 @@
-from modules import extra_networks
+from modules import extra_networks, shared, extra_networks
 from modules.hypernetworks import hypernetwork
 
 
@@ -7,6 +7,12 @@ class ExtraNetworkHypernet(extra_networks.ExtraNetwork):
         super().__init__('hypernet')
 
     def activate(self, p, params_list):
+        additional = shared.opts.sd_hypernetwork
+
+        if additional != "" and additional in shared.hypernetworks and len([x for x in params_list if x.items[0] == additional]) == 0:
+            p.all_prompts = [x + f"<hypernet:{additional}:{shared.opts.extra_networks_default_multiplier}>" for x in p.all_prompts]
+            params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier]))
+
         names = []
         multipliers = []
         for params in params_list:

modules/extras.py

@@ -6,7 +6,7 @@ import shutil
 import torch
 import tqdm
 
-from modules import shared, images, sd_models, sd_vae
+from modules import shared, images, sd_models, sd_vae, sd_models_config
 from modules.ui_common import plaintext_to_html
 import gradio as gr
 import safetensors.torch
@@ -37,7 +37,7 @@ def run_pnginfo(image):
 
 def create_config(ckpt_result, config_source, a, b, c):
     def config(x):
-        res = sd_models.find_checkpoint_config(x) if x else None
+        res = sd_models_config.find_checkpoint_config_near_filename(x) if x else None
         return res if res != shared.sd_default_config else None
 
     if config_source == 0:
@@ -132,6 +132,7 @@ def run_modelmerger(id_task, primary_model_name, secondary_model_name, tertiary_
     tertiary_model_info = sd_models.checkpoints_list[tertiary_model_name] if theta_func1 else None
 
     result_is_inpainting_model = False
+    result_is_instruct_pix2pix_model = False
 
     if theta_func2:
         shared.state.textinfo = f"Loading B"
@@ -185,14 +186,19 @@ def run_modelmerger(id_task, primary_model_name, secondary_model_name, tertiary_
             if a.shape != b.shape and a.shape[0:1] + a.shape[2:] == b.shape[0:1] + b.shape[2:]:
                 if a.shape[1] == 4 and b.shape[1] == 9:
                     raise RuntimeError("When merging inpainting model with a normal one, A must be the inpainting model.")
+                if a.shape[1] == 4 and b.shape[1] == 8:
+                    raise RuntimeError("When merging instruct-pix2pix model with a normal one, A must be the instruct-pix2pix model.")
 
-                assert a.shape[1] == 9 and b.shape[1] == 4, f"Bad dimensions for merged layer {key}: A={a.shape}, B={b.shape}"
-
-                theta_0[key][:, 0:4, :, :] = theta_func2(a[:, 0:4, :, :], b, multiplier)
-                result_is_inpainting_model = True
+                if a.shape[1] == 8 and b.shape[1] == 4:#If we have an Instruct-Pix2Pix model...
+                    theta_0[key][:, 0:4, :, :] = theta_func2(a[:, 0:4, :, :], b, multiplier)#Merge only the vectors the models have in common.  Otherwise we get an error due to dimension mismatch.
+                    result_is_instruct_pix2pix_model = True
+                else:
+                    assert a.shape[1] == 9 and b.shape[1] == 4, f"Bad dimensions for merged layer {key}: A={a.shape}, B={b.shape}"
+                    theta_0[key][:, 0:4, :, :] = theta_func2(a[:, 0:4, :, :], b, multiplier)
+                    result_is_inpainting_model = True
             else:
                 theta_0[key] = theta_func2(a, b, multiplier)
-
+            
             theta_0[key] = to_half(theta_0[key], save_as_half)
 
         shared.state.sampling_step += 1
@@ -226,6 +232,7 @@ def run_modelmerger(id_task, primary_model_name, secondary_model_name, tertiary_
 
     filename = filename_generator() if custom_name == '' else custom_name
     filename += ".inpainting" if result_is_inpainting_model else ""
+    filename += ".instruct-pix2pix" if result_is_instruct_pix2pix_model else ""
     filename += "." + checkpoint_format
 
     output_modelname = os.path.join(ckpt_dir, filename)

modules/generation_parameters_copypaste.py

@@ -6,14 +6,13 @@ import re
 from pathlib import Path
 
 import gradio as gr
-from modules.shared import script_path
+from modules.paths import data_path
 from modules import shared, ui_tempdir, script_callbacks
 import tempfile
 from PIL import Image
 
-re_param_code = r'\s*([\w ]+):\s*("(?:\\|\"|[^\"])+"|[^,]*)(?:,|$)'
+re_param_code = r'\s*([\w ]+):\s*("(?:\\"[^,]|\\"|\\|[^\"])+"|[^,]*)(?:,|$)'
 re_param = re.compile(re_param_code)
-re_params = re.compile(r"^(?:" + re_param_code + "){3,}$")
 re_imagesize = re.compile(r"^(\d+)x(\d+)$")
 re_hypernet_hash = re.compile("\(([0-9a-f]+)\)$")
 type_of_gr_update = type(gr.update())
@@ -243,7 +242,7 @@ Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 965400086, Size: 512x512, Model
     done_with_prompt = False
 
     *lines, lastline = x.strip().split("\n")
-    if not re_params.match(lastline):
+    if len(re_param.findall(lastline)) < 3:
         lines.append(lastline)
         lastline = ''
 
@@ -261,6 +260,7 @@ Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 965400086, Size: 512x512, Model
     res["Negative prompt"] = negative_prompt
 
     for k, v in re_param.findall(lastline):
+        v = v[1:-1] if v[0] == '"' and v[-1] == '"' else v
         m = re_imagesize.match(v)
         if m is not None:
             res[k+"-1"] = m.group(1)
@@ -293,7 +293,7 @@ Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 965400086, Size: 512x512, Model
 def connect_paste(button, paste_fields, input_comp, jsfunc=None):
     def paste_func(prompt):
         if not prompt and not shared.cmd_opts.hide_ui_dir_config:
-            filename = os.path.join(script_path, "params.txt")
+            filename = os.path.join(data_path, "params.txt")
             if os.path.exists(filename):
                 with open(filename, "r", encoding="utf8") as file:
                     prompt = file.read()

modules/gfpgan_model.py

@@ -6,12 +6,11 @@ import facexlib
 import gfpgan
 
 import modules.face_restoration
-from modules import shared, devices, modelloader
-from modules.paths import models_path
+from modules import paths, shared, devices, modelloader
 
 model_dir = "GFPGAN"
 user_path = None
-model_path = os.path.join(models_path, model_dir)
+model_path = os.path.join(paths.models_path, model_dir)
 model_url = "https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth"
 have_gfpgan = False
 loaded_gfpgan_model = None

modules/hashes.py

@@ -4,8 +4,10 @@ import os.path
 
 import filelock
 
+from modules.paths import data_path
 
-cache_filename = "cache.json"
+
+cache_filename = os.path.join(data_path, "cache.json")
 cache_data = None
 
 

modules/images.py

@@ -36,6 +36,8 @@ def image_grid(imgs, batch_size=1, rows=None):
         else:
             rows = math.sqrt(len(imgs))
             rows = round(rows)
+    if rows > len(imgs):
+        rows = len(imgs)
 
     cols = math.ceil(len(imgs) / rows)
 
@@ -195,7 +197,7 @@ def draw_grid_annotations(im, width, height, hor_texts, ver_texts):
     ver_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing * len(lines) for lines in
                         ver_texts]
 
-    pad_top = max(hor_text_heights) + line_spacing * 2
+    pad_top = 0 if sum(hor_text_heights) == 0 else max(hor_text_heights) + line_spacing * 2
 
     result = Image.new("RGB", (im.width + pad_left, im.height + pad_top), "white")
     result.paste(im, (pad_left, pad_top))

modules/img2img.py

@@ -16,11 +16,18 @@ import modules.images as images
 import modules.scripts
 
 
-def process_batch(p, input_dir, output_dir, args):
+def process_batch(p, input_dir, output_dir, inpaint_mask_dir, args):
     processing.fix_seed(p)
 
     images = shared.listfiles(input_dir)
 
+    is_inpaint_batch = False
+    if inpaint_mask_dir:
+        inpaint_masks = shared.listfiles(inpaint_mask_dir)
+        is_inpaint_batch = len(inpaint_masks) > 0
+    if is_inpaint_batch:
+        print(f"\nInpaint batch is enabled. {len(inpaint_masks)} masks found.")
+
     print(f"Will process {len(images)} images, creating {p.n_iter * p.batch_size} new images for each.")
 
     save_normally = output_dir == ''
@@ -43,6 +50,15 @@ def process_batch(p, input_dir, output_dir, args):
         img = ImageOps.exif_transpose(img)
         p.init_images = [img] * p.batch_size
 
+        if is_inpaint_batch:
+            # try to find corresponding mask for an image using simple filename matching
+            mask_image_path = os.path.join(inpaint_mask_dir, os.path.basename(image))
+            # if not found use first one ("same mask for all images" use-case)
+            if not mask_image_path in inpaint_masks:
+                mask_image_path = inpaint_masks[0]
+            mask_image = Image.open(mask_image_path)
+            p.image_mask = mask_image
+
         proc = modules.scripts.scripts_img2img.run(p, *args)
         if proc is None:
             proc = process_images(p)
@@ -59,7 +75,7 @@ def process_batch(p, input_dir, output_dir, args):
                 processed_image.save(os.path.join(output_dir, filename))
 
 
-def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps: int, sampler_index: int, mask_blur: int, mask_alpha: float, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, *args):
+def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps: int, sampler_index: int, mask_blur: int, mask_alpha: float, inpainting_fill: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, denoising_strength: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, img2img_batch_inpaint_mask_dir: str, *args):
     is_batch = mode == 5
 
     if mode == 0:  # img2img
@@ -139,7 +155,7 @@ def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_s
     if is_batch:
         assert not shared.cmd_opts.hide_ui_dir_config, "Launched with --hide-ui-dir-config, batch img2img disabled"
 
-        process_batch(p, img2img_batch_input_dir, img2img_batch_output_dir, args)
+        process_batch(p, img2img_batch_input_dir, img2img_batch_output_dir, img2img_batch_inpaint_mask_dir, args)
 
         processed = Processed(p, [], p.seed, "")
     else:

modules/interrogate.py

@@ -12,7 +12,7 @@ from torchvision import transforms
 from torchvision.transforms.functional import InterpolationMode
 
 import modules.shared as shared
-from modules import devices, paths, lowvram, modelloader, errors
+from modules import devices, paths, shared, lowvram, modelloader, errors
 
 blip_image_eval_size = 384
 clip_model_name = 'ViT-L/14'
@@ -82,9 +82,16 @@ class InterrogateModels:
 
         return self.loaded_categories
 
+    def create_fake_fairscale(self):
+        class FakeFairscale:
+            def checkpoint_wrapper(self):
+                pass
+
+        sys.modules["fairscale.nn.checkpoint.checkpoint_activations"] = FakeFairscale
+
     def load_blip_model(self):
-        with paths.Prioritize("BLIP"):
-            import models.blip
+        self.create_fake_fairscale()
+        import models.blip
 
         files = modelloader.load_models(
             model_path=os.path.join(paths.models_path, "BLIP"),

modules/paths.py

@@ -4,7 +4,15 @@ import sys
 import modules.safe
 
 script_path = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
-models_path = os.path.join(script_path, "models")
+
+# Parse the --data-dir flag first so we can use it as a base for our other argument default values
+parser = argparse.ArgumentParser(add_help=False)
+parser.add_argument("--data-dir", type=str, default=os.path.dirname(os.path.dirname(os.path.realpath(__file__))), help="base path where all user data is stored",)
+cmd_opts_pre = parser.parse_known_args()[0]
+data_path = cmd_opts_pre.data_dir
+models_path = os.path.join(data_path, "models")
+
+# data_path = cmd_opts_pre.data
 sys.path.insert(0, script_path)
 
 # search for directory of stable diffusion in following places

modules/processing.py

@@ -13,10 +13,11 @@ from skimage import exposure
 from typing import Any, Dict, List, Optional
 
 import modules.sd_hijack
-from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, script_callbacks, extra_networks
+from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, script_callbacks, extra_networks, sd_vae_approx, scripts
 from modules.sd_hijack import model_hijack
 from modules.shared import opts, cmd_opts, state
 import modules.shared as shared
+import modules.paths as paths
 import modules.face_restoration
 import modules.images as images
 import modules.styles
@@ -184,7 +185,12 @@ class StableDiffusionProcessing:
         conditioning = 2. * (conditioning - depth_min) / (depth_max - depth_min) - 1.
         return conditioning
 
-    def inpainting_image_conditioning(self, source_image, latent_image, image_mask = None):
+    def edit_image_conditioning(self, source_image):
+        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image))
+
+        return conditioning_image
+
+    def inpainting_image_conditioning(self, source_image, latent_image, image_mask=None):
         self.is_using_inpainting_conditioning = True
 
         # Handle the different mask inputs
@@ -203,7 +209,7 @@ class StableDiffusionProcessing:
 
         # Create another latent image, this time with a masked version of the original input.
         # Smoothly interpolate between the masked and unmasked latent conditioning image using a parameter.
-        conditioning_mask = conditioning_mask.to(source_image.device).to(source_image.dtype)
+        conditioning_mask = conditioning_mask.to(device=source_image.device, dtype=source_image.dtype)
         conditioning_image = torch.lerp(
             source_image,
             source_image * (1.0 - conditioning_mask),
@@ -222,11 +228,16 @@ class StableDiffusionProcessing:
         return image_conditioning
 
     def img2img_image_conditioning(self, source_image, latent_image, image_mask=None):
+        source_image = devices.cond_cast_float(source_image)
+
         # HACK: Using introspection as the Depth2Image model doesn't appear to uniquely
         # identify itself with a field common to all models. The conditioning_key is also hybrid.
         if isinstance(self.sd_model, LatentDepth2ImageDiffusion):
             return self.depth2img_image_conditioning(source_image)
 
+        if self.sd_model.cond_stage_key == "edit":
+            return self.edit_image_conditioning(source_image)
+
         if self.sampler.conditioning_key in {'hybrid', 'concat'}:
             return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
 
@@ -439,8 +450,6 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments=None, iter
         "Size": f"{p.width}x{p.height}",
         "Model hash": getattr(p, 'sd_model_hash', None if not opts.add_model_hash_to_info or not shared.sd_model.sd_model_hash else shared.sd_model.sd_model_hash),
         "Model": (None if not opts.add_model_name_to_info or not shared.sd_model.sd_checkpoint_info.model_name else shared.sd_model.sd_checkpoint_info.model_name.replace(',', '').replace(':', '')),
-        "Batch size": (None if p.batch_size < 2 else p.batch_size),
-        "Batch pos": (None if p.batch_size < 2 else position_in_batch),
         "Variation seed": (None if p.subseed_strength == 0 else all_subseeds[index]),
         "Variation seed strength": (None if p.subseed_strength == 0 else p.subseed_strength),
         "Seed resize from": (None if p.seed_resize_from_w == 0 or p.seed_resize_from_h == 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}"),
@@ -580,10 +589,14 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
         with devices.autocast():
             p.init(p.all_prompts, p.all_seeds, p.all_subseeds)
 
+            # for OSX, loading the model during sampling changes the generated picture, so it is loaded here
+            if shared.opts.live_previews_enable and opts.show_progress_type == "Approx NN":
+                sd_vae_approx.model()
+
             if not p.disable_extra_networks:
                 extra_networks.activate(p, extra_network_data)
 
-        with open(os.path.join(shared.script_path, "params.txt"), "w", encoding="utf8") as file:
+        with open(os.path.join(paths.data_path, "params.txt"), "w", encoding="utf8") as file:
             processed = Processed(p, [], p.seed, "")
             file.write(processed.infotext(p, 0))
 
@@ -634,7 +647,8 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
             if p.n_iter > 1:
                 shared.state.job = f"Batch {n+1} out of {p.n_iter}"
 
-            with devices.autocast():
+
+            with devices.without_autocast() if devices.unet_needs_upcast else devices.autocast():
                 if type(p) == StableDiffusionProcessingTxt2Img:
                     if p.enable_hr:
                         if p.hr_prompt != '':
@@ -684,6 +698,11 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
 
                 image = Image.fromarray(x_sample)
 
+                if p.scripts is not None:
+                    pp = scripts.PostprocessImageArgs(image)
+                    p.scripts.postprocess_image(p, pp)
+                    image = pp.image
+
                 if p.color_corrections is not None and i < len(p.color_corrections):
                     if opts.save and not p.do_not_save_samples and opts.save_images_before_color_correction:
                         image_without_cc = apply_overlay(image, p.paste_to, i, p.overlay_images)

modules/realesrgan_model.py

@@ -46,7 +46,7 @@ class UpscalerRealESRGAN(Upscaler):
             scale=info.scale,
             model_path=info.local_data_path,
             model=info.model(),
-            half=not cmd_opts.no_half,
+            half=not cmd_opts.no_half and not cmd_opts.upcast_sampling,
             tile=opts.ESRGAN_tile,
             tile_pad=opts.ESRGAN_tile_overlap,
         )

modules/script_loading.py

@@ -1,16 +1,14 @@
 import os
 import sys
 import traceback
+import importlib.util
 from types import ModuleType
 
 
 def load_module(path):
-    with open(path, "r", encoding="utf8") as file:
-        text = file.read()
-
-    compiled = compile(text, path, 'exec')
-    module = ModuleType(os.path.basename(path))
-    exec(compiled, module.__dict__)
+    module_spec = importlib.util.spec_from_file_location(os.path.basename(path), path)
+    module = importlib.util.module_from_spec(module_spec)
+    module_spec.loader.exec_module(module)
 
     return module
 

modules/scripts.py

@@ -6,12 +6,16 @@ from collections import namedtuple
 
 import gradio as gr
 
-from modules.processing import StableDiffusionProcessing
 from modules import shared, paths, script_callbacks, extensions, script_loading, scripts_postprocessing
 
 AlwaysVisible = object()
 
 
+class PostprocessImageArgs:
+    def __init__(self, image):
+        self.image = image
+
+
 class Script:
     filename = None
     args_from = None
@@ -65,7 +69,7 @@ class Script:
         args contains all values returned by components from ui()
         """
 
-        raise NotImplementedError()
+        pass
 
     def process(self, p, *args):
         """
@@ -100,6 +104,13 @@ class Script:
 
         pass
 
+    def postprocess_image(self, p, pp: PostprocessImageArgs, *args):
+        """
+        Called for every image after it has been generated.
+        """
+
+        pass
+
     def postprocess(self, p, processed, *args):
         """
         This function is called after processing ends for AlwaysVisible scripts.
@@ -247,11 +258,15 @@ class ScriptRunner:
         self.infotext_fields = []
 
     def initialize_scripts(self, is_img2img):
+        from modules import scripts_auto_postprocessing
+
         self.scripts.clear()
         self.alwayson_scripts.clear()
         self.selectable_scripts.clear()
 
-        for script_class, path, basedir, script_module in scripts_data:
+        auto_processing_scripts = scripts_auto_postprocessing.create_auto_preprocessing_script_data()
+
+        for script_class, path, basedir, script_module in auto_processing_scripts + scripts_data:
             script = script_class()
             script.filename = path
             script.is_txt2img = not is_img2img
@@ -330,9 +345,23 @@ class ScriptRunner:
             outputs=[script.group for script in self.selectable_scripts]
         )
 
+        self.script_load_ctr = 0
+        def onload_script_visibility(params):
+            title = params.get('Script', None)
+            if title:
+                title_index = self.titles.index(title)
+                visibility = title_index == self.script_load_ctr
+                self.script_load_ctr = (self.script_load_ctr + 1) % len(self.titles)
+                return gr.update(visible=visibility)
+            else:
+                return gr.update(visible=False)
+
+        self.infotext_fields.append( (dropdown, lambda x: gr.update(value=x.get('Script', 'None'))) )
+        self.infotext_fields.extend( [(script.group, onload_script_visibility) for script in self.selectable_scripts] )
+
         return inputs
 
-    def run(self, p: StableDiffusionProcessing, *args):
+    def run(self, p, *args):
         script_index = args[0]
 
         if script_index == 0:
@@ -386,6 +415,15 @@ class ScriptRunner:
                 print(f"Error running postprocess_batch: {script.filename}", file=sys.stderr)
                 print(traceback.format_exc(), file=sys.stderr)
 
+    def postprocess_image(self, p, pp: PostprocessImageArgs):
+        for script in self.alwayson_scripts:
+            try:
+                script_args = p.script_args[script.args_from:script.args_to]
+                script.postprocess_image(p, pp, *script_args)
+            except Exception:
+                print(f"Error running postprocess_batch: {script.filename}", file=sys.stderr)
+                print(traceback.format_exc(), file=sys.stderr)
+
     def before_component(self, component, **kwargs):
         for script in self.scripts:
             try:

modules/scripts_auto_postprocessing.py

@@ -0,0 +1,42 @@
+from modules import scripts, scripts_postprocessing, shared
+
+
+class ScriptPostprocessingForMainUI(scripts.Script):
+    def __init__(self, script_postproc):
+        self.script: scripts_postprocessing.ScriptPostprocessing = script_postproc
+        self.postprocessing_controls = None
+
+    def title(self):
+        return self.script.name
+
+    def show(self, is_img2img):
+        return scripts.AlwaysVisible
+
+    def ui(self, is_img2img):
+        self.postprocessing_controls = self.script.ui()
+        return self.postprocessing_controls.values()
+
+    def postprocess_image(self, p, script_pp, *args):
+        args_dict = {k: v for k, v in zip(self.postprocessing_controls, args)}
+
+        pp = scripts_postprocessing.PostprocessedImage(script_pp.image)
+        pp.info = {}
+        self.script.process(pp, **args_dict)
+        p.extra_generation_params.update(pp.info)
+        script_pp.image = pp.image
+
+
+def create_auto_preprocessing_script_data():
+    from modules import scripts
+
+    res = []
+
+    for name in shared.opts.postprocessing_enable_in_main_ui:
+        script = next(iter([x for x in scripts.postprocessing_scripts_data if x.script_class.name == name]), None)
+        if script is None:
+            continue
+
+        constructor = lambda s=script: ScriptPostprocessingForMainUI(s.script_class())
+        res.append(scripts.ScriptClassData(script_class=constructor, path=script.path, basedir=script.basedir, module=script.module))
+
+    return res

modules/scripts_postprocessing.py

@@ -46,6 +46,8 @@ class ScriptPostprocessing:
         pass
 
 
+
+
 def wrap_call(func, filename, funcname, *args, default=None, **kwargs):
     try:
         res = func(*args, **kwargs)
@@ -68,6 +70,9 @@ class ScriptPostprocessingRunner:
             script: ScriptPostprocessing = script_class()
             script.filename = path
 
+            if script.name == "Simple Upscale":
+                continue
+
             self.scripts.append(script)
 
     def create_script_ui(self, script, inputs):
@@ -87,12 +92,11 @@ class ScriptPostprocessingRunner:
             import modules.scripts
             self.initialize_scripts(modules.scripts.postprocessing_scripts_data)
 
-        scripts_order = [x.lower().strip() for x in shared.opts.postprocessing_scipts_order.split(",")]
+        scripts_order = shared.opts.postprocessing_operation_order
 
         def script_score(name):
-            name = name.lower()
             for i, possible_match in enumerate(scripts_order):
-                if possible_match in name:
+                if possible_match == name:
                     return i
 
             return len(self.scripts)
@@ -145,3 +149,4 @@ class ScriptPostprocessingRunner:
     def image_changed(self):
         for script in self.scripts_in_preferred_order():
             script.image_changed()
+

modules/sd_hijack.py

@@ -131,6 +131,8 @@ class StableDiffusionModelHijack:
             m.cond_stage_model.wrapped.model.token_embedding = m.cond_stage_model.wrapped.model.token_embedding.wrapped
             m.cond_stage_model = m.cond_stage_model.wrapped
 
+        undo_optimizations()
+
         self.apply_circular(False)
         self.layers = None
         self.clip = None
@@ -171,7 +173,7 @@ class EmbeddingsWithFixes(torch.nn.Module):
         vecs = []
         for fixes, tensor in zip(batch_fixes, inputs_embeds):
             for offset, embedding in fixes:
-                emb = embedding.vec
+                emb = devices.cond_cast_unet(embedding.vec)
                 emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0])
                 tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]])
 

modules/sd_hijack_inpainting.py

@@ -96,15 +96,6 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F
     return x_prev, pred_x0, e_t
 
 
-def should_hijack_inpainting(checkpoint_info):
-    from modules import sd_models
-
-    ckpt_basename = os.path.basename(checkpoint_info.filename).lower()
-    cfg_basename = os.path.basename(sd_models.find_checkpoint_config(checkpoint_info)).lower()
-
-    return "inpainting" in ckpt_basename and not "inpainting" in cfg_basename
-
-
 def do_inpainting_hijack():
     # p_sample_plms is needed because PLMS can't work with dicts as conditionings
 

modules/sd_hijack_ip2p.py

@@ -0,0 +1,13 @@
+import collections
+import os.path
+import sys
+import gc
+import time
+
+def should_hijack_ip2p(checkpoint_info):
+    from modules import sd_models_config
+
+    ckpt_basename = os.path.basename(checkpoint_info.filename).lower()
+    cfg_basename = os.path.basename(sd_models_config.find_checkpoint_config_near_filename(checkpoint_info)).lower()
+
+    return "pix2pix" in ckpt_basename and not "pix2pix" in cfg_basename

modules/sd_hijack_optimizations.py

@@ -9,7 +9,7 @@ from torch import einsum
 from ldm.util import default
 from einops import rearrange
 
-from modules import shared, errors
+from modules import shared, errors, devices
 from modules.hypernetworks import hypernetwork
 
 from .sub_quadratic_attention import efficient_dot_product_attention
@@ -52,18 +52,25 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None):
     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
     del q_in, k_in, v_in
 
-    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device)
-    for i in range(0, q.shape[0], 2):
-        end = i + 2
-        s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end])
-        s1 *= self.scale
+    dtype = q.dtype
+    if shared.opts.upcast_attn:
+        q, k, v = q.float(), k.float(), v.float()
 
-        s2 = s1.softmax(dim=-1)
-        del s1
+    with devices.without_autocast(disable=not shared.opts.upcast_attn):
+        r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+        for i in range(0, q.shape[0], 2):
+            end = i + 2
+            s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end])
+            s1 *= self.scale
+    
+            s2 = s1.softmax(dim=-1)
+            del s1
+    
+            r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end])
+            del s2
+        del q, k, v
 
-        r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end])
-        del s2
-    del q, k, v
+    r1 = r1.to(dtype)
 
     r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
     del r1
@@ -82,45 +89,52 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
     k_in = self.to_k(context_k)
     v_in = self.to_v(context_v)
 
-    k_in *= self.scale
+    dtype = q_in.dtype
+    if shared.opts.upcast_attn:
+        q_in, k_in, v_in = q_in.float(), k_in.float(), v_in if v_in.device.type == 'mps' else v_in.float()
 
-    del context, x
+    with devices.without_autocast(disable=not shared.opts.upcast_attn):
+        k_in = k_in * self.scale
+    
+        del context, x
+    
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
+        del q_in, k_in, v_in
+    
+        r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+    
+        mem_free_total = get_available_vram()
+    
+        gb = 1024 ** 3
+        tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size()
+        modifier = 3 if q.element_size() == 2 else 2.5
+        mem_required = tensor_size * modifier
+        steps = 1
+    
+        if mem_required > mem_free_total:
+            steps = 2 ** (math.ceil(math.log(mem_required / mem_free_total, 2)))
+            # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
+            #       f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
+    
+        if steps > 64:
+            max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
+            raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
+                               f'Need: {mem_required / 64 / gb:0.1f}GB free, Have:{mem_free_total / gb:0.1f}GB free')
+    
+        slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
+        for i in range(0, q.shape[1], slice_size):
+            end = i + slice_size
+            s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k)
+    
+            s2 = s1.softmax(dim=-1, dtype=q.dtype)
+            del s1
+    
+            r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
+            del s2
+    
+        del q, k, v
 
-    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
-    del q_in, k_in, v_in
-
-    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
-
-    mem_free_total = get_available_vram()
-
-    gb = 1024 ** 3
-    tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size()
-    modifier = 3 if q.element_size() == 2 else 2.5
-    mem_required = tensor_size * modifier
-    steps = 1
-
-    if mem_required > mem_free_total:
-        steps = 2 ** (math.ceil(math.log(mem_required / mem_free_total, 2)))
-        # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
-        #       f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
-
-    if steps > 64:
-        max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
-        raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
-                           f'Need: {mem_required / 64 / gb:0.1f}GB free, Have:{mem_free_total / gb:0.1f}GB free')
-
-    slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
-    for i in range(0, q.shape[1], slice_size):
-        end = i + slice_size
-        s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k)
-
-        s2 = s1.softmax(dim=-1, dtype=q.dtype)
-        del s1
-
-        r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
-        del s2
-
-    del q, k, v
+    r1 = r1.to(dtype)
 
     r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
     del r1
@@ -204,12 +218,20 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
     context = default(context, x)
 
     context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
-    k = self.to_k(context_k) * self.scale
+    k = self.to_k(context_k)
     v = self.to_v(context_v)
     del context, context_k, context_v, x
 
-    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
-    r = einsum_op(q, k, v)
+    dtype = q.dtype
+    if shared.opts.upcast_attn:
+        q, k, v = q.float(), k.float(), v if v.device.type == 'mps' else v.float()
+
+    with devices.without_autocast(disable=not shared.opts.upcast_attn):
+        k = k * self.scale
+    
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        r = einsum_op(q, k, v)
+    r = r.to(dtype)
     return self.to_out(rearrange(r, '(b h) n d -> b n (h d)', h=h))
 
 # -- End of code from https://github.com/invoke-ai/InvokeAI --
@@ -234,8 +256,14 @@ def sub_quad_attention_forward(self, x, context=None, mask=None):
     k = k.unflatten(-1, (h, -1)).transpose(1,2).flatten(end_dim=1)
     v = v.unflatten(-1, (h, -1)).transpose(1,2).flatten(end_dim=1)
 
+    dtype = q.dtype
+    if shared.opts.upcast_attn:
+        q, k = q.float(), k.float()
+
     x = sub_quad_attention(q, k, v, q_chunk_size=shared.cmd_opts.sub_quad_q_chunk_size, kv_chunk_size=shared.cmd_opts.sub_quad_kv_chunk_size, chunk_threshold=shared.cmd_opts.sub_quad_chunk_threshold, use_checkpoint=self.training)
 
+    x = x.to(dtype)
+
     x = x.unflatten(0, (-1, h)).transpose(1,2).flatten(start_dim=2)
 
     out_proj, dropout = self.to_out
@@ -268,15 +296,16 @@ def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_
         query_chunk_size = q_tokens
         kv_chunk_size = k_tokens
 
-    return efficient_dot_product_attention(
-        q,
-        k,
-        v,
-        query_chunk_size=q_chunk_size,
-        kv_chunk_size=kv_chunk_size,
-        kv_chunk_size_min = kv_chunk_size_min,
-        use_checkpoint=use_checkpoint,
-    )
+    with devices.without_autocast(disable=q.dtype == v.dtype):
+        return efficient_dot_product_attention(
+            q,
+            k,
+            v,
+            query_chunk_size=q_chunk_size,
+            kv_chunk_size=kv_chunk_size,
+            kv_chunk_size_min = kv_chunk_size_min,
+            use_checkpoint=use_checkpoint,
+        )
 
 
 def get_xformers_flash_attention_op(q, k, v):
@@ -306,8 +335,14 @@ def xformers_attention_forward(self, x, context=None, mask=None):
     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q_in, k_in, v_in))
     del q_in, k_in, v_in
 
+    dtype = q.dtype
+    if shared.opts.upcast_attn:
+        q, k = q.float(), k.float()
+
     out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=get_xformers_flash_attention_op(q, k, v))
 
+    out = out.to(dtype)
+
     out = rearrange(out, 'b n h d -> b n (h d)', h=h)
     return self.to_out(out)
 
@@ -378,10 +413,14 @@ def xformers_attnblock_forward(self, x):
         v = self.v(h_)
         b, c, h, w = q.shape
         q, k, v = map(lambda t: rearrange(t, 'b c h w -> b (h w) c'), (q, k, v))
+        dtype = q.dtype
+        if shared.opts.upcast_attn:
+            q, k = q.float(), k.float()
         q = q.contiguous()
         k = k.contiguous()
         v = v.contiguous()
         out = xformers.ops.memory_efficient_attention(q, k, v, op=get_xformers_flash_attention_op(q, k, v))
+        out = out.to(dtype)
         out = rearrange(out, 'b (h w) c -> b c h w', h=h)
         out = self.proj_out(out)
         return x + out

modules/sd_hijack_unet.py

@@ -1,4 +1,8 @@
 import torch
+from packaging import version
+
+from modules import devices
+from modules.sd_hijack_utils import CondFunc
 
 
 class TorchHijackForUnet:
@@ -28,3 +32,37 @@ class TorchHijackForUnet:
 
 
 th = TorchHijackForUnet()
+
+
+# Below are monkey patches to enable upcasting a float16 UNet for float32 sampling
+def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
+
+    if isinstance(cond, dict):
+        for y in cond.keys():
+            cond[y] = [x.to(devices.dtype_unet) if isinstance(x, torch.Tensor) else x for x in cond[y]]
+
+    with devices.autocast():
+        return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
+
+class GELUHijack(torch.nn.GELU, torch.nn.Module):
+    def __init__(self, *args, **kwargs):
+        torch.nn.GELU.__init__(self, *args, **kwargs)
+    def forward(self, x):
+        if devices.unet_needs_upcast:
+            return torch.nn.GELU.forward(self.float(), x.float()).to(devices.dtype_unet)
+        else:
+            return torch.nn.GELU.forward(self, x)
+
+unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
+CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)
+if version.parse(torch.__version__) <= version.parse("1.13.1"):
+    CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)
+    CondFunc('ldm.modules.attention.GEGLU.forward', lambda orig_func, self, x: orig_func(self.float(), x.float()).to(devices.dtype_unet), unet_needs_upcast)
+    CondFunc('open_clip.transformer.ResidualAttentionBlock.__init__', lambda orig_func, *args, **kwargs: kwargs.update({'act_layer': GELUHijack}) and False or orig_func(*args, **kwargs), lambda _, *args, **kwargs: kwargs.get('act_layer') is None or kwargs['act_layer'] == torch.nn.GELU)
+
+first_stage_cond = lambda _, self, *args, **kwargs: devices.unet_needs_upcast and self.model.diffusion_model.dtype == torch.float16
+first_stage_sub = lambda orig_func, self, x, **kwargs: orig_func(self, x.to(devices.dtype_vae), **kwargs)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)

modules/sd_hijack_utils.py

@@ -0,0 +1,28 @@
+import importlib
+
+class CondFunc:
+    def __new__(cls, orig_func, sub_func, cond_func):
+        self = super(CondFunc, cls).__new__(cls)
+        if isinstance(orig_func, str):
+            func_path = orig_func.split('.')
+            for i in range(len(func_path)-1, -1, -1):
+                try:
+                    resolved_obj = importlib.import_module('.'.join(func_path[:i]))
+                    break
+                except ImportError:
+                    pass
+            for attr_name in func_path[i:-1]:
+                resolved_obj = getattr(resolved_obj, attr_name)
+            orig_func = getattr(resolved_obj, func_path[-1])
+            setattr(resolved_obj, func_path[-1], lambda *args, **kwargs: self(*args, **kwargs))
+        self.__init__(orig_func, sub_func, cond_func)
+        return lambda *args, **kwargs: self(*args, **kwargs)
+    def __init__(self, orig_func, sub_func, cond_func):
+        self.__orig_func = orig_func
+        self.__sub_func = sub_func
+        self.__cond_func = cond_func
+    def __call__(self, *args, **kwargs):
+        if not self.__cond_func or self.__cond_func(self.__orig_func, *args, **kwargs):
+            return self.__sub_func(self.__orig_func, *args, **kwargs)
+        else:
+            return self.__orig_func(*args, **kwargs)

modules/sd_models.py

@@ -2,8 +2,6 @@ import collections
 import os.path
 import sys
 import gc
-import time
-from collections import namedtuple
 import torch
 import re
 import safetensors.torch
@@ -14,12 +12,13 @@ import ldm.modules.midas as midas
 
 from ldm.util import instantiate_from_config
 
-from modules import shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors, hashes
+from modules import paths, shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors, hashes, sd_models_config
 from modules.paths import models_path
-from modules.sd_hijack_inpainting import do_inpainting_hijack, should_hijack_inpainting
+from modules.sd_hijack_inpainting import do_inpainting_hijack
+from modules.timer import Timer
 
 model_dir = "Stable-diffusion"
-model_path = os.path.abspath(os.path.join(models_path, model_dir))
+model_path = os.path.abspath(os.path.join(paths.models_path, model_dir))
 
 checkpoints_list = {}
 checkpoint_alisases = {}
@@ -42,6 +41,7 @@ class CheckpointInfo:
             name = name[1:]
 
         self.name = name
+        self.name_for_extra = os.path.splitext(os.path.basename(filename))[0]
         self.model_name = os.path.splitext(name.replace("/", "_").replace("\\", "_"))[0]
         self.hash = model_hash(filename)
 
@@ -98,17 +98,6 @@ def checkpoint_tiles():
     return sorted([x.title for x in checkpoints_list.values()], key=alphanumeric_key)
 
 
-def find_checkpoint_config(info):
-    if info is None:
-        return shared.cmd_opts.config
-
-    config = os.path.splitext(info.filename)[0] + ".yaml"
-    if os.path.exists(config):
-        return config
-
-    return shared.cmd_opts.config
-
-
 def list_models():
     checkpoints_list.clear()
     checkpoint_alisases.clear()
@@ -214,9 +203,7 @@ def get_state_dict_from_checkpoint(pl_sd):
 def read_state_dict(checkpoint_file, print_global_state=False, map_location=None):
     _, extension = os.path.splitext(checkpoint_file)
     if extension.lower() == ".safetensors":
-        device = map_location or shared.weight_load_location
-        if device is None:
-            device = devices.get_cuda_device_string() if torch.cuda.is_available() else "cpu"
+        device = map_location or shared.weight_load_location or devices.get_optimal_device_name()
         pl_sd = safetensors.torch.load_file(checkpoint_file, device=device)
     else:
         pl_sd = torch.load(checkpoint_file, map_location=map_location or shared.weight_load_location)
@@ -228,52 +215,72 @@ def read_state_dict(checkpoint_file, print_global_state=False, map_location=None
     return sd
 
 
-def load_model_weights(model, checkpoint_info: CheckpointInfo):
-    title = checkpoint_info.title
+def get_checkpoint_state_dict(checkpoint_info: CheckpointInfo, timer):
     sd_model_hash = checkpoint_info.calculate_shorthash()
-    if checkpoint_info.title != title:
-        shared.opts.data["sd_model_checkpoint"] = checkpoint_info.title
+    timer.record("calculate hash")
 
-    cache_enabled = shared.opts.sd_checkpoint_cache > 0
-
-    if cache_enabled and checkpoint_info in checkpoints_loaded:
+    if checkpoint_info in checkpoints_loaded:
         # use checkpoint cache
         print(f"Loading weights [{sd_model_hash}] from cache")
-        model.load_state_dict(checkpoints_loaded[checkpoint_info])
-    else:
-        # load from file
-        print(f"Loading weights [{sd_model_hash}] from {checkpoint_info.filename}")
+        return checkpoints_loaded[checkpoint_info]
 
-        sd = read_state_dict(checkpoint_info.filename)
-        model.load_state_dict(sd, strict=False)
-        del sd
-        
-        if cache_enabled:
-            # cache newly loaded model
-            checkpoints_loaded[checkpoint_info] = model.state_dict().copy()
+    print(f"Loading weights [{sd_model_hash}] from {checkpoint_info.filename}")
+    res = read_state_dict(checkpoint_info.filename)
+    timer.record("load weights from disk")
 
-        if shared.cmd_opts.opt_channelslast:
-            model.to(memory_format=torch.channels_last)
+    return res
 
-        if not shared.cmd_opts.no_half:
-            vae = model.first_stage_model
 
-            # with --no-half-vae, remove VAE from model when doing half() to prevent its weights from being converted to float16
-            if shared.cmd_opts.no_half_vae:
-                model.first_stage_model = None
+def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer):
+    sd_model_hash = checkpoint_info.calculate_shorthash()
+    timer.record("calculate hash")
 
-            model.half()
-            model.first_stage_model = vae
+    shared.opts.data["sd_model_checkpoint"] = checkpoint_info.title
 
-        devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16
-        devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16
+    if state_dict is None:
+        state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
 
-        model.first_stage_model.to(devices.dtype_vae)
+    model.load_state_dict(state_dict, strict=False)
+    del state_dict
+    timer.record("apply weights to model")
+
+    if shared.opts.sd_checkpoint_cache > 0:
+        # cache newly loaded model
+        checkpoints_loaded[checkpoint_info] = model.state_dict().copy()
+
+    if shared.cmd_opts.opt_channelslast:
+        model.to(memory_format=torch.channels_last)
+        timer.record("apply channels_last")
+
+    if not shared.cmd_opts.no_half:
+        vae = model.first_stage_model
+        depth_model = getattr(model, 'depth_model', None)
+
+        # with --no-half-vae, remove VAE from model when doing half() to prevent its weights from being converted to float16
+        if shared.cmd_opts.no_half_vae:
+            model.first_stage_model = None
+        # with --upcast-sampling, don't convert the depth model weights to float16
+        if shared.cmd_opts.upcast_sampling and depth_model:
+            model.depth_model = None
+
+        model.half()
+        model.first_stage_model = vae
+        if depth_model:
+            model.depth_model = depth_model
+
+        timer.record("apply half()")
+
+    devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16
+    devices.dtype_vae = torch.float32 if shared.cmd_opts.no_half or shared.cmd_opts.no_half_vae else torch.float16
+    devices.dtype_unet = model.model.diffusion_model.dtype
+    devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16
+
+    model.first_stage_model.to(devices.dtype_vae)
+    timer.record("apply dtype to VAE")
 
     # clean up cache if limit is reached
-    if cache_enabled:
-        while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache + 1: # we need to count the current model
-            checkpoints_loaded.popitem(last=False)  # LRU
+    while len(checkpoints_loaded) > shared.opts.sd_checkpoint_cache:
+        checkpoints_loaded.popitem(last=False)
 
     model.sd_model_hash = sd_model_hash
     model.sd_model_checkpoint = checkpoint_info.filename
@@ -286,6 +293,7 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo):
     sd_vae.clear_loaded_vae()
     vae_file, vae_source = sd_vae.resolve_vae(checkpoint_info.filename)
     sd_vae.load_vae(model, vae_file, vae_source)
+    timer.record("load VAE")
 
 
 def enable_midas_autodownload():
@@ -298,7 +306,7 @@ def enable_midas_autodownload():
     location automatically.
     """
 
-    midas_path = os.path.join(models_path, 'midas')
+    midas_path = os.path.join(paths.models_path, 'midas')
 
     # stable-diffusion-stability-ai hard-codes the midas model path to
     # a location that differs from where other scripts using this model look.
@@ -331,24 +339,20 @@ def enable_midas_autodownload():
     midas.api.load_model = load_model_wrapper
 
 
-class Timer:
-    def __init__(self):
-        self.start = time.time()
+def repair_config(sd_config):
 
-    def elapsed(self):
-        end = time.time()
-        res = end - self.start
-        self.start = end
-        return res
+    if not hasattr(sd_config.model.params, "use_ema"):
+        sd_config.model.params.use_ema = False
+
+    if shared.cmd_opts.no_half:
+        sd_config.model.params.unet_config.params.use_fp16 = False
+    elif shared.cmd_opts.upcast_sampling:
+        sd_config.model.params.unet_config.params.use_fp16 = True
 
 
-def load_model(checkpoint_info=None):
+def load_model(checkpoint_info=None, already_loaded_state_dict=None, time_taken_to_load_state_dict=None):
     from modules import lowvram, sd_hijack
     checkpoint_info = checkpoint_info or select_checkpoint()
-    checkpoint_config = find_checkpoint_config(checkpoint_info)
-
-    if checkpoint_config != shared.cmd_opts.config:
-        print(f"Loading config from: {checkpoint_config}")
 
     if shared.sd_model:
         sd_hijack.model_hijack.undo_hijack(shared.sd_model)
@@ -356,27 +360,27 @@ def load_model(checkpoint_info=None):
         gc.collect()
         devices.torch_gc()
 
-    sd_config = OmegaConf.load(checkpoint_config)
-    
-    if should_hijack_inpainting(checkpoint_info):
-        # Hardcoded config for now...
-        sd_config.model.target = "ldm.models.diffusion.ddpm.LatentInpaintDiffusion"
-        sd_config.model.params.conditioning_key = "hybrid"
-        sd_config.model.params.unet_config.params.in_channels = 9
-        sd_config.model.params.finetune_keys = None
-
-    if not hasattr(sd_config.model.params, "use_ema"):
-        sd_config.model.params.use_ema = False
-
     do_inpainting_hijack()
 
-    if shared.cmd_opts.no_half:
-        sd_config.model.params.unet_config.params.use_fp16 = False
-
     timer = Timer()
 
-    sd_model = None
+    if already_loaded_state_dict is not None:
+        state_dict = already_loaded_state_dict
+    else:
+        state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
 
+    checkpoint_config = sd_models_config.find_checkpoint_config(state_dict, checkpoint_info)
+
+    timer.record("find config")
+
+    sd_config = OmegaConf.load(checkpoint_config)
+    repair_config(sd_config)
+
+    timer.record("load config")
+
+    print(f"Creating model from config: {checkpoint_config}")
+
+    sd_model = None
     try:
         with sd_disable_initialization.DisableInitialization():
             sd_model = instantiate_from_config(sd_config.model)
@@ -387,29 +391,35 @@ def load_model(checkpoint_info=None):
         print('Failed to create model quickly; will retry using slow method.', file=sys.stderr)
         sd_model = instantiate_from_config(sd_config.model)
 
-    elapsed_create = timer.elapsed()
+    sd_model.used_config = checkpoint_config
 
-    load_model_weights(sd_model, checkpoint_info)
+    timer.record("create model")
 
-    elapsed_load_weights = timer.elapsed()
+    load_model_weights(sd_model, checkpoint_info, state_dict, timer)
 
     if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
         lowvram.setup_for_low_vram(sd_model, shared.cmd_opts.medvram)
     else:
         sd_model.to(shared.device)
 
+    timer.record("move model to device")
+
     sd_hijack.model_hijack.hijack(sd_model)
 
+    timer.record("hijack")
+
     sd_model.eval()
     shared.sd_model = sd_model
 
     sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings(force_reload=True)  # Reload embeddings after model load as they may or may not fit the model
 
+    timer.record("load textual inversion embeddings")
+
     script_callbacks.model_loaded_callback(sd_model)
 
-    elapsed_the_rest = timer.elapsed()
+    timer.record("scripts callbacks")
 
-    print(f"Model loaded in {elapsed_create + elapsed_load_weights + elapsed_the_rest:.1f}s ({elapsed_create:.1f}s create model, {elapsed_load_weights:.1f}s load weights).")
+    print(f"Model loaded in {timer.summary()}.")
 
     return sd_model
 
@@ -420,6 +430,7 @@ def reload_model_weights(sd_model=None, info=None):
 
     if not sd_model:
         sd_model = shared.sd_model
+
     if sd_model is None:  # previous model load failed
         current_checkpoint_info = None
     else:
@@ -427,38 +438,44 @@ def reload_model_weights(sd_model=None, info=None):
         if sd_model.sd_model_checkpoint == checkpoint_info.filename:
             return
 
-    checkpoint_config = find_checkpoint_config(current_checkpoint_info)
+        if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
+            lowvram.send_everything_to_cpu()
+        else:
+            sd_model.to(devices.cpu)
 
-    if current_checkpoint_info is None or checkpoint_config != find_checkpoint_config(checkpoint_info) or should_hijack_inpainting(checkpoint_info) != should_hijack_inpainting(sd_model.sd_checkpoint_info):
-        del sd_model
-        checkpoints_loaded.clear()
-        load_model(checkpoint_info)
-        return shared.sd_model
-
-    if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
-        lowvram.send_everything_to_cpu()
-    else:
-        sd_model.to(devices.cpu)
-
-    sd_hijack.model_hijack.undo_hijack(sd_model)
+        sd_hijack.model_hijack.undo_hijack(sd_model)
 
     timer = Timer()
 
+    state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
+
+    checkpoint_config = sd_models_config.find_checkpoint_config(state_dict, checkpoint_info)
+
+    timer.record("find config")
+
+    if sd_model is None or checkpoint_config != sd_model.used_config:
+        del sd_model
+        checkpoints_loaded.clear()
+        load_model(checkpoint_info, already_loaded_state_dict=state_dict, time_taken_to_load_state_dict=timer.records["load weights from disk"])
+        return shared.sd_model
+
     try:
-        load_model_weights(sd_model, checkpoint_info)
+        load_model_weights(sd_model, checkpoint_info, state_dict, timer)
     except Exception as e:
         print("Failed to load checkpoint, restoring previous")
-        load_model_weights(sd_model, current_checkpoint_info)
+        load_model_weights(sd_model, current_checkpoint_info, None, timer)
         raise
     finally:
         sd_hijack.model_hijack.hijack(sd_model)
+        timer.record("hijack")
+
         script_callbacks.model_loaded_callback(sd_model)
+        timer.record("script callbacks")
 
         if not shared.cmd_opts.lowvram and not shared.cmd_opts.medvram:
             sd_model.to(devices.device)
+            timer.record("move model to device")
 
-    elapsed = timer.elapsed()
-
-    print(f"Weights loaded in {elapsed:.1f}s.")
+    print(f"Weights loaded in {timer.summary()}.")
 
     return sd_model

modules/sd_models_config.py

@@ -0,0 +1,112 @@
+import re
+import os
+
+import torch
+
+from modules import shared, paths, sd_disable_initialization
+
+sd_configs_path = shared.sd_configs_path
+sd_repo_configs_path = os.path.join(paths.paths['Stable Diffusion'], "configs", "stable-diffusion")
+
+
+config_default = shared.sd_default_config
+config_sd2 = os.path.join(sd_repo_configs_path, "v2-inference.yaml")
+config_sd2v = os.path.join(sd_repo_configs_path, "v2-inference-v.yaml")
+config_sd2_inpainting = os.path.join(sd_repo_configs_path, "v2-inpainting-inference.yaml")
+config_depth_model = os.path.join(sd_repo_configs_path, "v2-midas-inference.yaml")
+config_inpainting = os.path.join(sd_configs_path, "v1-inpainting-inference.yaml")
+config_instruct_pix2pix = os.path.join(sd_configs_path, "instruct-pix2pix.yaml")
+config_alt_diffusion = os.path.join(sd_configs_path, "alt-diffusion-inference.yaml")
+
+
+def is_using_v_parameterization_for_sd2(state_dict):
+    """
+    Detects whether unet in state_dict is using v-parameterization. Returns True if it is. You're welcome.
+    """
+
+    import ldm.modules.diffusionmodules.openaimodel
+    from modules import devices
+
+    device = devices.cpu
+
+    with sd_disable_initialization.DisableInitialization():
+        unet = ldm.modules.diffusionmodules.openaimodel.UNetModel(
+            use_checkpoint=True,
+            use_fp16=False,
+            image_size=32,
+            in_channels=4,
+            out_channels=4,
+            model_channels=320,
+            attention_resolutions=[4, 2, 1],
+            num_res_blocks=2,
+            channel_mult=[1, 2, 4, 4],
+            num_head_channels=64,
+            use_spatial_transformer=True,
+            use_linear_in_transformer=True,
+            transformer_depth=1,
+            context_dim=1024,
+            legacy=False
+        )
+        unet.eval()
+
+    with torch.no_grad():
+        unet_sd = {k.replace("model.diffusion_model.", ""): v for k, v in state_dict.items() if "model.diffusion_model." in k}
+        unet.load_state_dict(unet_sd, strict=True)
+        unet.to(device=device, dtype=torch.float)
+
+        test_cond = torch.ones((1, 2, 1024), device=device) * 0.5
+        x_test = torch.ones((1, 4, 8, 8), device=device) * 0.5
+
+        out = (unet(x_test, torch.asarray([999], device=device), context=test_cond) - x_test).mean().item()
+
+    return out < -1
+
+
+def guess_model_config_from_state_dict(sd, filename):
+    sd2_cond_proj_weight = sd.get('cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight', None)
+    diffusion_model_input = sd.get('model.diffusion_model.input_blocks.0.0.weight', None)
+
+    if sd.get('depth_model.model.pretrained.act_postprocess3.0.project.0.bias', None) is not None:
+        return config_depth_model
+
+    if sd2_cond_proj_weight is not None and sd2_cond_proj_weight.shape[1] == 1024:
+        if diffusion_model_input.shape[1] == 9:
+            return config_sd2_inpainting
+        elif is_using_v_parameterization_for_sd2(sd):
+            return config_sd2v
+        else:
+            return config_sd2
+
+    if diffusion_model_input is not None:
+        if diffusion_model_input.shape[1] == 9:
+            return config_inpainting
+        if diffusion_model_input.shape[1] == 8:
+            return config_instruct_pix2pix
+
+    if sd.get('cond_stage_model.roberta.embeddings.word_embeddings.weight', None) is not None:
+        return config_alt_diffusion
+
+    return config_default
+
+
+def find_checkpoint_config(state_dict, info):
+    if info is None:
+        return guess_model_config_from_state_dict(state_dict, "")
+
+    config = find_checkpoint_config_near_filename(info)
+    if config is not None:
+        return config
+
+    return guess_model_config_from_state_dict(state_dict, info.filename)
+
+
+def find_checkpoint_config_near_filename(info):
+    if info is None:
+        return None
+
+    config = os.path.splitext(info.filename)[0] + ".yaml"
+    if os.path.exists(config):
+        return config
+
+    return None
+

modules/sd_samplers.py

@@ -454,7 +454,7 @@ class KDiffusionSampler:
     def initialize(self, p):
         self.model_wrap_cfg.mask = p.mask if hasattr(p, 'mask') else None
         self.model_wrap_cfg.nmask = p.nmask if hasattr(p, 'nmask') else None
-        self.model_wrap.step = 0
+        self.model_wrap_cfg.step = 0
         self.eta = p.eta or opts.eta_ancestral
 
         k_diffusion.sampling.torch = TorchHijack(self.sampler_noises if self.sampler_noises is not None else [])

modules/sd_vae.py

@@ -3,13 +3,12 @@ import safetensors.torch
 import os
 import collections
 from collections import namedtuple
-from modules import shared, devices, script_callbacks, sd_models
-from modules.paths import models_path
+from modules import paths, shared, devices, script_callbacks, sd_models
 import glob
 from copy import deepcopy
 
 
-vae_path = os.path.abspath(os.path.join(models_path, "VAE"))
+vae_path = os.path.abspath(os.path.join(paths.models_path, "VAE"))
 vae_ignore_keys = {"model_ema.decay", "model_ema.num_updates"}
 vae_dict = {}
 

modules/shared.py

@@ -13,17 +13,19 @@ import modules.interrogate
 import modules.memmon
 import modules.styles
 import modules.devices as devices
-from modules import localization, sd_vae, extensions, script_loading, errors, ui_components
-from modules.paths import models_path, script_path, sd_path
+from modules import localization, extensions, script_loading, errors, ui_components, shared_items
+from modules.paths import models_path, script_path, data_path
 
 
 demo = None
 
-sd_default_config = os.path.join(script_path, "configs/v1-inference.yaml")
+sd_configs_path = os.path.join(script_path, "configs")
+sd_default_config = os.path.join(sd_configs_path, "v1-inference.yaml")
 sd_model_file = os.path.join(script_path, 'model.ckpt')
 default_sd_model_file = sd_model_file
 
 parser = argparse.ArgumentParser()
+parser.add_argument("--data-dir", type=str, default=os.path.dirname(os.path.dirname(os.path.realpath(__file__))), help="base path where all user data is stored",)
 parser.add_argument("--config", type=str, default=sd_default_config, help="path to config which constructs model",)
 parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; if specified, this checkpoint will be added to the list of checkpoints and loaded",)
 parser.add_argument("--ckpt-dir", type=str, default=None, help="Path to directory with stable diffusion checkpoints")
@@ -34,7 +36,7 @@ parser.add_argument("--no-half", action='store_true', help="do not switch the mo
 parser.add_argument("--no-half-vae", action='store_true', help="do not switch the VAE model to 16-bit floats")
 parser.add_argument("--no-progressbar-hiding", action='store_true', help="do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware acceleration in browser)")
 parser.add_argument("--max-batch-count", type=int, default=16, help="maximum batch count value for the UI")
-parser.add_argument("--embeddings-dir", type=str, default=os.path.join(script_path, 'embeddings'), help="embeddings directory for textual inversion (default: embeddings)")
+parser.add_argument("--embeddings-dir", type=str, default=os.path.join(data_path, 'embeddings'), help="embeddings directory for textual inversion (default: embeddings)")
 parser.add_argument("--textual-inversion-templates-dir", type=str, default=os.path.join(script_path, 'textual_inversion_templates'), help="directory with textual inversion templates")
 parser.add_argument("--hypernetwork-dir", type=str, default=os.path.join(models_path, 'hypernetworks'), help="hypernetwork directory")
 parser.add_argument("--localizations-dir", type=str, default=os.path.join(script_path, 'localizations'), help="localizations directory")
@@ -45,6 +47,7 @@ parser.add_argument("--lowram", action='store_true', help="load stable diffusion
 parser.add_argument("--always-batch-cond-uncond", action='store_true', help="disables cond/uncond batching that is enabled to save memory with --medvram or --lowvram")
 parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.")
 parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast")
+parser.add_argument("--upcast-sampling", action='store_true', help="upcast sampling. No effect with --no-half. Usually produces similar results to --no-half with better performance while using less memory.")
 parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site")
 parser.add_argument("--ngrok", type=str, help="ngrok authtoken, alternative to gradio --share", default=None)
 parser.add_argument("--ngrok-region", type=str, help="The region in which ngrok should start.", default="us")
@@ -72,16 +75,16 @@ parser.add_argument("--use-cpu", nargs='+', help="use CPU as torch device for sp
 parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests")
 parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None)
 parser.add_argument("--show-negative-prompt", action='store_true', help="does not do anything", default=False)
-parser.add_argument("--ui-config-file", type=str, help="filename to use for ui configuration", default=os.path.join(script_path, 'ui-config.json'))
+parser.add_argument("--ui-config-file", type=str, help="filename to use for ui configuration", default=os.path.join(data_path, 'ui-config.json'))
 parser.add_argument("--hide-ui-dir-config", action='store_true', help="hide directory configuration from webui", default=False)
 parser.add_argument("--freeze-settings", action='store_true', help="disable editing settings", default=False)
-parser.add_argument("--ui-settings-file", type=str, help="filename to use for ui settings", default=os.path.join(script_path, 'config.json'))
+parser.add_argument("--ui-settings-file", type=str, help="filename to use for ui settings", default=os.path.join(data_path, 'config.json'))
 parser.add_argument("--gradio-debug",  action='store_true', help="launch gradio with --debug option")
 parser.add_argument("--gradio-auth", type=str, help='set gradio authentication like "username:password"; or comma-delimit multiple like "u1:p1,u2:p2,u3:p3"', default=None)
 parser.add_argument("--gradio-img2img-tool", type=str, help='does not do anything')
 parser.add_argument("--gradio-inpaint-tool", type=str, help="does not do anything")
 parser.add_argument("--opt-channelslast", action='store_true', help="change memory type for stable diffusion to channels last")
-parser.add_argument("--styles-file", type=str, help="filename to use for styles", default=os.path.join(script_path, 'styles.csv'))
+parser.add_argument("--styles-file", type=str, help="filename to use for styles", default=os.path.join(data_path, 'styles.csv'))
 parser.add_argument("--autolaunch", action='store_true', help="open the webui URL in the system's default browser upon launch", default=False)
 parser.add_argument("--theme", type=str, help="launches the UI with light or dark theme", default=None)
 parser.add_argument("--use-textbox-seed", action='store_true', help="use textbox for seeds in UI (no up/down, but possible to input long seeds)", default=False)
@@ -263,12 +266,6 @@ interrogator = modules.interrogate.InterrogateModels("interrogate")
 
 face_restorers = []
 
-
-def realesrgan_models_names():
-    import modules.realesrgan_model
-    return [x.name for x in modules.realesrgan_model.get_realesrgan_models(None)]
-
-
 class OptionInfo:
     def __init__(self, default=None, label="", component=None, component_args=None, onchange=None, section=None, refresh=None):
         self.default = default
@@ -359,7 +356,7 @@ options_templates.update(options_section(('saving-to-dirs', "Saving to a directo
 options_templates.update(options_section(('upscaling', "Upscaling"), {
     "ESRGAN_tile": OptionInfo(192, "Tile size for ESRGAN upscalers. 0 = no tiling.", gr.Slider, {"minimum": 0, "maximum": 512, "step": 16}),
     "ESRGAN_tile_overlap": OptionInfo(8, "Tile overlap, in pixels for ESRGAN upscalers. Low values = visible seam.", gr.Slider, {"minimum": 0, "maximum": 48, "step": 1}),
-    "realesrgan_enabled_models": OptionInfo(["R-ESRGAN 4x+", "R-ESRGAN 4x+ Anime6B"], "Select which Real-ESRGAN models to show in the web UI. (Requires restart)", gr.CheckboxGroup, lambda: {"choices": realesrgan_models_names()}),
+    "realesrgan_enabled_models": OptionInfo(["R-ESRGAN 4x+", "R-ESRGAN 4x+ Anime6B"], "Select which Real-ESRGAN models to show in the web UI. (Requires restart)", gr.CheckboxGroup, lambda: {"choices": shared_items.realesrgan_models_names()}),
     "upscaler_for_img2img": OptionInfo(None, "Upscaler for img2img", gr.Dropdown, lambda: {"choices": [x.name for x in sd_upscalers]}),
 }))
 
@@ -396,7 +393,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": list_checkpoint_tiles()}, refresh=refresh_checkpoints),
     "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}),
     "sd_vae_checkpoint_cache": OptionInfo(0, "VAE Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}),
-    "sd_vae": OptionInfo("Automatic", "SD VAE", gr.Dropdown, lambda: {"choices": ["Automatic", "None"] + list(sd_vae.vae_dict)}, refresh=sd_vae.refresh_vae_list),
+    "sd_vae": OptionInfo("Automatic", "SD VAE", gr.Dropdown, lambda: {"choices": shared_items.sd_vae_items()}, refresh=shared_items.refresh_vae_list),
     "sd_vae_as_default": OptionInfo(True, "Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them"),
     "inpainting_mask_weight": OptionInfo(1.0, "Inpainting conditioning mask strength", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     "initial_noise_multiplier": OptionInfo(1.0, "Noise multiplier for img2img", gr.Slider, {"minimum": 0.5, "maximum": 1.5, "step": 0.01}),
@@ -408,7 +405,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
     "comma_padding_backtrack": OptionInfo(20, "Increase coherency by padding from the last comma within n tokens when using more than 75 tokens", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1 }),
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}),
-    "extra_networks_default_multiplier": OptionInfo(1.0, "Multiplier for extra networks", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
+    "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
 }))
 
 options_templates.update(options_section(('compatibility', "Compatibility"), {
@@ -433,7 +430,9 @@ options_templates.update(options_section(('interrogate', "Interrogate Options"),
 }))
 
 options_templates.update(options_section(('extra_networks', "Extra Networks"), {
-    "extra_networks_default_view": OptionInfo("cards", "Default view for Extra Networks", gr.Dropdown, { "choices": ["cards", "thumbs"] }),
+    "extra_networks_default_view": OptionInfo("cards", "Default view for Extra Networks", gr.Dropdown, {"choices": ["cards", "thumbs"]}),
+    "extra_networks_default_multiplier": OptionInfo(1.0, "Multiplier for extra networks", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
+    "sd_hypernetwork": OptionInfo("None", "Add hypernetwork to prompt", gr.Dropdown, lambda: {"choices": [""] + [x for x in hypernetworks.keys()]}, refresh=reload_hypernetworks),
 }))
 
 options_templates.update(options_section(('ui', "User interface"), {
@@ -481,7 +480,8 @@ options_templates.update(options_section(('sampler-params', "Sampler parameters"
 }))
 
 options_templates.update(options_section(('postprocessing', "Postprocessing"), {
-    'postprocessing_scipts_order': OptionInfo("upscale, gfpgan, codeformer", "Postprocessing operation order"),
+    'postprocessing_enable_in_main_ui': OptionInfo([], "Enable postprocessing operations in txt2img and img2img tabs", ui_components.DropdownMulti, lambda: {"choices": [x.name for x in shared_items.postprocessing_scripts()]}),
+    'postprocessing_operation_order': OptionInfo([], "Postprocessing operation order", ui_components.DropdownMulti, lambda: {"choices": [x.name for x in shared_items.postprocessing_scripts()]}),
     'upscaling_max_images_in_cache': OptionInfo(5, "Maximum number of images in upscaling cache", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}),
 }))
 

modules/shared_items.py

@@ -0,0 +1,23 @@
+
+
+def realesrgan_models_names():
+    import modules.realesrgan_model
+    return [x.name for x in modules.realesrgan_model.get_realesrgan_models(None)]
+
+
+def postprocessing_scripts():
+    import modules.scripts
+
+    return modules.scripts.scripts_postproc.scripts
+
+
+def sd_vae_items():
+    import modules.sd_vae
+
+    return ["Automatic", "None"] + list(modules.sd_vae.vae_dict)
+
+
+def refresh_vae_list():
+    import modules.sd_vae
+
+    return modules.sd_vae.refresh_vae_list

modules/sub_quadratic_attention.py

@@ -67,7 +67,7 @@ def _summarize_chunk(
     max_score, _ = torch.max(attn_weights, -1, keepdim=True)
     max_score = max_score.detach()
     exp_weights = torch.exp(attn_weights - max_score)
-    exp_values = torch.bmm(exp_weights, value)
+    exp_values = torch.bmm(exp_weights, value) if query.device.type == 'mps' else torch.bmm(exp_weights, value.to(exp_weights.dtype)).to(value.dtype)
     max_score = max_score.squeeze(-1)
     return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score)
 
@@ -129,7 +129,7 @@ def _get_attention_scores_no_kv_chunking(
     )
     attn_probs = attn_scores.softmax(dim=-1)
     del attn_scores
-    hidden_states_slice = torch.bmm(attn_probs, value)
+    hidden_states_slice = torch.bmm(attn_probs, value) if query.device.type == 'mps' else torch.bmm(attn_probs, value.to(attn_probs.dtype)).to(value.dtype)
     return hidden_states_slice
 
 

modules/textual_inversion/preprocess.py

@@ -6,8 +6,7 @@ import sys
 import tqdm
 import time
 
-from modules import shared, images, deepbooru
-from modules.paths import models_path
+from modules import paths, shared, images, deepbooru
 from modules.shared import opts, cmd_opts
 from modules.textual_inversion import autocrop
 
@@ -199,7 +198,7 @@ def preprocess_work(process_src, process_dst, process_width, process_height, pre
 
             dnn_model_path = None
             try:
-                dnn_model_path = autocrop.download_and_cache_models(os.path.join(models_path, "opencv"))
+                dnn_model_path = autocrop.download_and_cache_models(os.path.join(paths.models_path, "opencv"))
             except Exception as e:
                 print("Unable to load face detection model for auto crop selection. Falling back to lower quality haar method.", e)
 

modules/textual_inversion/textual_inversion.py

@@ -112,6 +112,7 @@ class EmbeddingDatabase:
         self.skipped_embeddings = {}
         self.expected_shape = -1
         self.embedding_dirs = {}
+        self.previously_displayed_embeddings = ()
 
     def add_embedding_dir(self, path):
         self.embedding_dirs[path] = DirWithTextualInversionEmbeddings(path)
@@ -194,7 +195,7 @@ class EmbeddingDatabase:
         if not os.path.isdir(embdir.path):
             return
 
-        for root, dirs, fns in os.walk(embdir.path):
+        for root, dirs, fns in os.walk(embdir.path, followlinks=True):
             for fn in fns:
                 try:
                     fullfn = os.path.join(root, fn)
@@ -228,9 +229,12 @@ class EmbeddingDatabase:
             self.load_from_dir(embdir)
             embdir.update()
 
-        print(f"Textual inversion embeddings loaded({len(self.word_embeddings)}): {', '.join(self.word_embeddings.keys())}")
-        if len(self.skipped_embeddings) > 0:
-            print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings.keys())}")
+        displayed_embeddings = (tuple(self.word_embeddings.keys()), tuple(self.skipped_embeddings.keys()))
+        if self.previously_displayed_embeddings != displayed_embeddings:
+            self.previously_displayed_embeddings = displayed_embeddings
+            print(f"Textual inversion embeddings loaded({len(self.word_embeddings)}): {', '.join(self.word_embeddings.keys())}")
+            if len(self.skipped_embeddings) > 0:
+                print(f"Textual inversion embeddings skipped({len(self.skipped_embeddings)}): {', '.join(self.skipped_embeddings.keys())}")
 
     def find_embedding_at_position(self, tokens, offset):
         token = tokens[offset]

modules/timer.py

@@ -0,0 +1,35 @@
+import time
+
+
+class Timer:
+    def __init__(self):
+        self.start = time.time()
+        self.records = {}
+        self.total = 0
+
+    def elapsed(self):
+        end = time.time()
+        res = end - self.start
+        self.start = end
+        return res
+
+    def record(self, category, extra_time=0):
+        e = self.elapsed()
+        if category not in self.records:
+            self.records[category] = 0
+
+        self.records[category] += e + extra_time
+        self.total += e + extra_time
+
+    def summary(self):
+        res = f"{self.total:.1f}s"
+
+        additions = [x for x in self.records.items() if x[1] >= 0.1]
+        if not additions:
+            return res
+
+        res += " ("
+        res += ", ".join([f"{category}: {time_taken:.1f}s" for category, time_taken in additions])
+        res += ")"
+
+        return res

modules/ui.py

@@ -21,7 +21,7 @@ from modules.call_queue import wrap_gradio_gpu_call, wrap_queued_call, wrap_grad
 
 from modules import sd_hijack, sd_models, localization, script_callbacks, ui_extensions, deepbooru, sd_vae, extra_networks, postprocessing, ui_components, ui_common, ui_postprocessing
 from modules.ui_components import FormRow, FormGroup, ToolButton, FormHTML
-from modules.paths import script_path
+from modules.paths import script_path, data_path
 
 from modules.shared import opts, cmd_opts, restricted_opts
 
@@ -91,6 +91,7 @@ save_style_symbol = '\U0001f4be'  # 💾
 apply_style_symbol = '\U0001f4cb'  # 📋
 clear_prompt_symbol = '\U0001F5D1'  # 🗑️
 extra_networks_symbol = '\U0001F3B4'  # 🎴
+switch_values_symbol = '\U000021C5' # ⇅
 
 
 def plaintext_to_html(text):
@@ -466,6 +467,7 @@ def create_ui():
                                 height = gr.Slider(minimum=64, maximum=2048, step=8, label="Height", value=512, elem_id="txt2img_height")
 
                             if opts.dimensions_and_batch_together:
+                                res_switch_btn = ToolButton(value=switch_values_symbol, elem_id="txt2img_res_switch_btn")
                                 with gr.Column(elem_id="txt2img_column_batch"):
                                     batch_count = gr.Slider(minimum=1, step=1, label='Batch count', value=1, elem_id="txt2img_batch_count")
                                     batch_size = gr.Slider(minimum=1, maximum=8, step=1, label='Batch size', value=1, elem_id="txt2img_batch_size")
@@ -581,6 +583,8 @@ def create_ui():
             txt2img_prompt.submit(**txt2img_args)
             submit.click(**txt2img_args)
 
+            res_switch_btn.click(lambda w, h: (h, w), inputs=[width, height], outputs=[width, height])
+
             txt_prompt_img.change(
                 fn=modules.images.image_data,
                 inputs=[
@@ -708,9 +712,15 @@ def create_ui():
 
                     with gr.TabItem('Batch', id='batch', elem_id="img2img_batch_tab") as tab_batch:
                         hidden = '<br>Disabled when launched with --hide-ui-dir-config.' if shared.cmd_opts.hide_ui_dir_config else ''
-                        gr.HTML(f"<p style='padding-bottom: 1em;' class=\"text-gray-500\">Process images in a directory on the same machine where the server is running.<br>Use an empty output directory to save pictures normally instead of writing to the output directory.{hidden}</p>")
+                        gr.HTML(
+                            f"<p style='padding-bottom: 1em;' class=\"text-gray-500\">Process images in a directory on the same machine where the server is running." +
+                            f"<br>Use an empty output directory to save pictures normally instead of writing to the output directory." +
+                            f"<br>Add inpaint batch mask directory to enable inpaint batch processing."
+                            f"{hidden}</p>"
+                        )
                         img2img_batch_input_dir = gr.Textbox(label="Input directory", **shared.hide_dirs, elem_id="img2img_batch_input_dir")
                         img2img_batch_output_dir = gr.Textbox(label="Output directory", **shared.hide_dirs, elem_id="img2img_batch_output_dir")
+                        img2img_batch_inpaint_mask_dir = gr.Textbox(label="Inpaint batch mask directory (required for inpaint batch processing only)", **shared.hide_dirs, elem_id="img2img_batch_inpaint_mask_dir")
 
                 def copy_image(img):
                     if isinstance(img, dict) and 'image' in img:
@@ -745,6 +755,7 @@ def create_ui():
                                 height = gr.Slider(minimum=64, maximum=2048, step=8, label="Height", value=512, elem_id="img2img_height")
 
                             if opts.dimensions_and_batch_together:
+                                res_switch_btn = ToolButton(value=switch_values_symbol, elem_id="img2img_res_switch_btn")
                                 with gr.Column(elem_id="img2img_column_batch"):
                                     batch_count = gr.Slider(minimum=1, step=1, label='Batch count', value=1, elem_id="img2img_batch_count")
                                     batch_size = gr.Slider(minimum=1, maximum=8, step=1, label='Batch size', value=1, elem_id="img2img_batch_size")
@@ -855,6 +866,7 @@ def create_ui():
                     inpainting_mask_invert,
                     img2img_batch_input_dir,
                     img2img_batch_output_dir,
+                    img2img_batch_inpaint_mask_dir
                 ] + custom_inputs,
                 outputs=[
                     img2img_gallery,
@@ -882,6 +894,7 @@ def create_ui():
 
             img2img_prompt.submit(**img2img_args)
             submit.click(**img2img_args)
+            res_switch_btn.click(lambda w, h: (h, w), inputs=[width, height], outputs=[width, height])
 
             img2img_interrogate.click(
                 fn=lambda *args: process_interrogate(interrogate, *args),
@@ -1514,8 +1527,8 @@ def create_ui():
         with open(cssfile, "r", encoding="utf8") as file:
             css += file.read() + "\n"
 
-    if os.path.exists(os.path.join(script_path, "user.css")):
-        with open(os.path.join(script_path, "user.css"), "r", encoding="utf8") as file:
+    if os.path.exists(os.path.join(data_path, "user.css")):
+        with open(os.path.join(data_path, "user.css"), "r", encoding="utf8") as file:
             css += file.read() + "\n"
 
     if not cmd_opts.no_progressbar_hiding:
@@ -1564,6 +1577,14 @@ def create_ui():
                 outputs=[component, text_settings],
             )
 
+        button_set_checkpoint = gr.Button('Change checkpoint', elem_id='change_checkpoint', visible=False)
+        button_set_checkpoint.click(
+            fn=lambda value, _: run_settings_single(value, key='sd_model_checkpoint'),
+            _js="function(v){ var res = desiredCheckpointName; desiredCheckpointName = ''; return [res || v, null]; }",
+            inputs=[component_dict['sd_model_checkpoint'], dummy_component],
+            outputs=[component_dict['sd_model_checkpoint'], text_settings],
+        )
+
         component_keys = [k for k in opts.data_labels.keys() if k in component_dict]
 
         def get_settings_values():
@@ -1696,14 +1717,14 @@ def create_ui():
 
 
 def reload_javascript():
-    head = f'<script type="text/javascript" src="file={os.path.abspath("script.js")}"></script>\n'
+    head = f'<script type="text/javascript" src="file={os.path.abspath("script.js")}?{os.path.getmtime("script.js")}"></script>\n'
 
     inline = f"{localization.localization_js(shared.opts.localization)};"
     if cmd_opts.theme is not None:
         inline += f"set_theme('{cmd_opts.theme}');"
 
     for script in modules.scripts.list_scripts("javascript", ".js"):
-        head += f'<script type="text/javascript" src="file={script.path}"></script>\n'
+        head += f'<script type="text/javascript" src="file={script.path}?{os.path.getmtime(script.path)}"></script>\n'
 
     head += f'<script type="text/javascript">{inline}</script>\n'
 

modules/ui_components.py

@@ -48,3 +48,11 @@ class FormColorPicker(gr.ColorPicker, gr.components.FormComponent):
     def get_block_name(self):
         return "colorpicker"
 
+
+class DropdownMulti(gr.Dropdown):
+    """Same as gr.Dropdown but always multiselect"""
+    def __init__(self, **kwargs):
+        super().__init__(multiselect=True, **kwargs)
+
+    def get_block_name(self):
+        return "dropdown"

modules/ui_extensions.py

@@ -13,7 +13,7 @@ import shutil
 import errno
 
 from modules import extensions, shared, paths
-
+from modules.call_queue import wrap_gradio_gpu_call
 
 available_extensions = {"extensions": []}
 
@@ -50,12 +50,17 @@ def apply_and_restart(disable_list, update_list):
     shared.state.need_restart = True
 
 
-def check_updates():
+def check_updates(id_task, disable_list):
     check_access()
 
-    for ext in extensions.extensions:
-        if ext.remote is None:
-            continue
+    disabled = json.loads(disable_list)
+    assert type(disabled) == list, f"wrong disable_list data for apply_and_restart: {disable_list}"
+
+    exts = [ext for ext in extensions.extensions if ext.remote is not None and ext.name not in disabled]
+    shared.state.job_count = len(exts)
+
+    for ext in exts:
+        shared.state.textinfo = ext.name
 
         try:
             ext.check_updates()
@@ -63,7 +68,9 @@ def check_updates():
             print(f"Error checking updates for {ext.name}:", file=sys.stderr)
             print(traceback.format_exc(), file=sys.stderr)
 
-    return extension_table()
+        shared.state.nextjob()
+
+    return extension_table(), ""
 
 
 def extension_table():
@@ -132,7 +139,7 @@ def install_extension_from_url(dirname, url):
     normalized_url = normalize_git_url(url)
     assert len([x for x in extensions.extensions if normalize_git_url(x.remote) == normalized_url]) == 0, 'Extension with this URL is already installed'
 
-    tmpdir = os.path.join(paths.script_path, "tmp", dirname)
+    tmpdir = os.path.join(paths.data_path, "tmp", dirname)
 
     try:
         shutil.rmtree(tmpdir, True)
@@ -273,12 +280,13 @@ def create_ui():
         with gr.Tabs(elem_id="tabs_extensions") as tabs:
             with gr.TabItem("Installed"):
 
-                with gr.Row():
+                with gr.Row(elem_id="extensions_installed_top"):
                     apply = gr.Button(value="Apply and restart UI", variant="primary")
                     check = gr.Button(value="Check for updates")
                     extensions_disabled_list = gr.Text(elem_id="extensions_disabled_list", visible=False).style(container=False)
                     extensions_update_list = gr.Text(elem_id="extensions_update_list", visible=False).style(container=False)
 
+                info = gr.HTML()
                 extensions_table = gr.HTML(lambda: extension_table())
 
                 apply.click(
@@ -289,10 +297,10 @@ def create_ui():
                 )
 
                 check.click(
-                    fn=check_updates,
+                    fn=wrap_gradio_gpu_call(check_updates, extra_outputs=[gr.update()]),
                     _js="extensions_check",
-                    inputs=[],
-                    outputs=[extensions_table],
+                    inputs=[info, extensions_disabled_list],
+                    outputs=[extensions_table, info],
                 )
 
             with gr.TabItem("Available"):

modules/ui_extra_networks.py

@@ -1,4 +1,7 @@
+import glob
 import os.path
+import urllib.parse
+from pathlib import Path
 
 from modules import shared
 import gradio as gr
@@ -8,12 +11,31 @@ import html
 from modules.generation_parameters_copypaste import image_from_url_text
 
 extra_pages = []
+allowed_dirs = set()
 
 
 def register_page(page):
     """registers extra networks page for the UI; recommend doing it in on_before_ui() callback for extensions"""
 
     extra_pages.append(page)
+    allowed_dirs.clear()
+    allowed_dirs.update(set(sum([x.allowed_directories_for_previews() for x in extra_pages], [])))
+
+
+def add_pages_to_demo(app):
+    def fetch_file(filename: str = ""):
+        from starlette.responses import FileResponse
+
+        if not any([Path(x).resolve() in Path(filename).resolve().parents for x in allowed_dirs]):
+            raise ValueError(f"File cannot be fetched: {filename}. Must be in one of directories registered by extra pages.")
+
+        if os.path.splitext(filename)[1].lower() != ".png":
+            raise ValueError(f"File cannot be fetched: {filename}. Only png.")
+
+        # would profit from returning 304
+        return FileResponse(filename, headers={"Accept-Ranges": "bytes"})
+
+    app.add_api_route("/sd_extra_networks/thumb", fetch_file, methods=["GET"])
 
 
 class ExtraNetworksPage:
@@ -26,10 +48,44 @@ class ExtraNetworksPage:
     def refresh(self):
         pass
 
+    def link_preview(self, filename):
+        return "./sd_extra_networks/thumb?filename=" + urllib.parse.quote(filename.replace('\\', '/')) + "&mtime=" + str(os.path.getmtime(filename))
+
+    def search_terms_from_path(self, filename, possible_directories=None):
+        abspath = os.path.abspath(filename)
+
+        for parentdir in (possible_directories if possible_directories is not None else self.allowed_directories_for_previews()):
+            parentdir = os.path.abspath(parentdir)
+            if abspath.startswith(parentdir):
+                return abspath[len(parentdir):].replace('\\', '/')
+
+        return ""
+
     def create_html(self, tabname):
         view = shared.opts.extra_networks_default_view
         items_html = ''
 
+        subdirs = {}
+        for parentdir in [os.path.abspath(x) for x in self.allowed_directories_for_previews()]:
+            for x in glob.glob(os.path.join(parentdir, '**/*'), recursive=True):
+                if not os.path.isdir(x):
+                    continue
+
+                subdir = os.path.abspath(x)[len(parentdir):].replace("\\", "/")
+                while subdir.startswith("/"):
+                    subdir = subdir[1:]
+
+                subdirs[subdir] = 1
+
+        if subdirs:
+            subdirs = {"": 1, **subdirs}
+
+        subdirs_html = "".join([f"""
+<button class='gr-button gr-button-lg gr-button-secondary{" search-all" if subdir=="" else ""}' onclick='extraNetworksSearchButton("{tabname}_extra_tabs", event)'>
+{html.escape(subdir if subdir!="" else "all")}
+</button>
+""" for subdir in subdirs])
+
         for item in self.list_items():
             items_html += self.create_html_for_item(item, tabname)
 
@@ -38,6 +94,9 @@ class ExtraNetworksPage:
             items_html = shared.html("extra-networks-no-cards.html").format(dirs=dirs)
 
         res = f"""
+<div id='{tabname}_{self.name}_subdirs' class='extra-network-subdirs extra-network-subdirs-{view}'>
+{subdirs_html}
+</div>
 <div id='{tabname}_{self.name}_cards' class='extra-network-{view}'>
 {items_html}
 </div>
@@ -54,14 +113,19 @@ class ExtraNetworksPage:
     def create_html_for_item(self, item, tabname):
         preview = item.get("preview", None)
 
+        onclick = item.get("onclick", None)
+        if onclick is None:
+            onclick = '"' + html.escape(f"""return cardClicked({json.dumps(tabname)}, {item["prompt"]}, {"true" if self.allow_negative_prompt else "false"})""") + '"'
+
         args = {
             "preview_html": "style='background-image: url(\"" + html.escape(preview) + "\")'" if preview else '',
-            "prompt": item["prompt"],
+            "prompt": item.get("prompt", None),
             "tabname": json.dumps(tabname),
             "local_preview": json.dumps(item["local_preview"]),
             "name": item["name"],
-            "card_clicked": '"' + html.escape(f"""return cardClicked({json.dumps(tabname)}, {item["prompt"]}, {"true" if self.allow_negative_prompt else "false"})""") + '"',
+            "card_clicked": onclick,
             "save_card_preview": '"' + html.escape(f"""return saveCardPreview(event, {json.dumps(tabname)}, {json.dumps(item["local_preview"])})""") + '"',
+            "search_term": item.get("search_term", ""),
         }
 
         return self.card_page.format(**args)
@@ -117,8 +181,13 @@ def create_ui(container, button, tabname):
     ui.button_save_preview = gr.Button('Save preview', elem_id=tabname+"_save_preview", visible=False)
     ui.preview_target_filename = gr.Textbox('Preview save filename', elem_id=tabname+"_preview_filename", visible=False)
 
-    button.click(fn=lambda: gr.update(visible=True), inputs=[], outputs=[container])
-    button_close.click(fn=lambda: gr.update(visible=False), inputs=[], outputs=[container])
+    def toggle_visibility(is_visible):
+        is_visible = not is_visible
+        return is_visible, gr.update(visible=is_visible)
+
+    state_visible = gr.State(value=False)
+    button.click(fn=toggle_visibility, inputs=[state_visible], outputs=[state_visible, container])
+    button_close.click(fn=toggle_visibility, inputs=[state_visible], outputs=[state_visible, container])
 
     def refresh():
         res = []
@@ -138,7 +207,7 @@ def path_is_parent(parent_path, child_path):
     parent_path = os.path.abspath(parent_path)
     child_path = os.path.abspath(child_path)
 
-    return os.path.commonpath([parent_path]) == os.path.commonpath([parent_path, child_path])
+    return child_path.startswith(parent_path)
 
 
 def setup_ui(ui, gallery):
@@ -168,7 +237,8 @@ def setup_ui(ui, gallery):
 
     ui.button_save_preview.click(
         fn=save_preview,
-        _js="function(x, y, z){console.log(x, y, z); return [selected_gallery_index(), y, z]}",
+        _js="function(x, y, z){return [selected_gallery_index(), y, z]}",
         inputs=[ui.preview_target_filename, gallery, ui.preview_target_filename],
         outputs=[*ui.pages]
     )
+

modules/ui_extra_networks_checkpoints.py

@@ -0,0 +1,38 @@
+import html
+import json
+import os
+import urllib.parse
+
+from modules import shared, ui_extra_networks, sd_models
+
+
+class ExtraNetworksPageCheckpoints(ui_extra_networks.ExtraNetworksPage):
+    def __init__(self):
+        super().__init__('Checkpoints')
+
+    def refresh(self):
+        shared.refresh_checkpoints()
+
+    def list_items(self):
+        for name, checkpoint in sd_models.checkpoints_list.items():
+            path, ext = os.path.splitext(checkpoint.filename)
+            previews = [path + ".png", path + ".preview.png"]
+
+            preview = None
+            for file in previews:
+                if os.path.isfile(file):
+                    preview = self.link_preview(file)
+                    break
+
+            yield {
+                "name": checkpoint.name_for_extra,
+                "filename": path,
+                "preview": preview,
+                "search_term": self.search_terms_from_path(checkpoint.filename),
+                "onclick": '"' + html.escape(f"""return selectCheckpoint({json.dumps(name)})""") + '"',
+                "local_preview": path + ".png",
+            }
+
+    def allowed_directories_for_previews(self):
+        return [v for v in [shared.cmd_opts.ckpt_dir, sd_models.model_path] if v is not None]
+

modules/ui_extra_networks_hypernets.py

@@ -19,13 +19,14 @@ class ExtraNetworksPageHypernetworks(ui_extra_networks.ExtraNetworksPage):
             preview = None
             for file in previews:
                 if os.path.isfile(file):
-                    preview = "./file=" + file.replace('\\', '/') + "?mtime=" + str(os.path.getmtime(file))
+                    preview = self.link_preview(file)
                     break
 
             yield {
                 "name": name,
                 "filename": path,
                 "preview": preview,
+                "search_term": self.search_terms_from_path(path),
                 "prompt": json.dumps(f"<hypernet:{name}:") + " + opts.extra_networks_default_multiplier + " + json.dumps(">"),
                 "local_preview": path + ".png",
             }

modules/ui_extra_networks_textual_inversion.py

@@ -19,12 +19,13 @@ class ExtraNetworksPageTextualInversion(ui_extra_networks.ExtraNetworksPage):
 
             preview = None
             if os.path.isfile(preview_file):
-                preview = "./file=" + preview_file.replace('\\', '/') + "?mtime=" + str(os.path.getmtime(preview_file))
+                preview = self.link_preview(preview_file)
 
             yield {
                 "name": embedding.name,
                 "filename": embedding.filename,
                 "preview": preview,
+                "search_term": self.search_terms_from_path(embedding.filename),
                 "prompt": json.dumps(embedding.name),
                 "local_preview": path + ".preview.png",
             }

modules/upscaler.py

@@ -11,7 +11,6 @@ from modules import modelloader, shared
 
 LANCZOS = (Image.Resampling.LANCZOS if hasattr(Image, 'Resampling') else Image.LANCZOS)
 NEAREST = (Image.Resampling.NEAREST if hasattr(Image, 'Resampling') else Image.NEAREST)
-from modules.paths import models_path
 
 
 class Upscaler:
@@ -39,7 +38,7 @@ class Upscaler:
         self.mod_scale = None
 
         if self.model_path is None and self.name:
-            self.model_path = os.path.join(models_path, self.name)
+            self.model_path = os.path.join(shared.models_path, self.name)
         if self.model_path and create_dirs:
             os.makedirs(self.model_path, exist_ok=True)
 
@@ -143,4 +142,4 @@ class UpscalerNearest(Upscaler):
     def __init__(self, dirname=None):
         super().__init__(False)
         self.name = "Nearest"
-        self.scalers = [UpscalerData("Nearest", None, self)]
+        self.scalers = [UpscalerData("Nearest", None, self)]