Merge branch 'master' of https://github.com/AUTOMATIC1111/stable-diffusion-webui into Base

scripts/img2imgalt.py

@@ -59,7 +59,7 @@ def find_noise_for_image(p, cond, uncond, cfg_scale, steps):
     return x / x.std()
 
 
-Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt"])
+Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt", "original_negative_prompt"])
 
 
 class Script(scripts.Script):
@@ -74,34 +74,45 @@ class Script(scripts.Script):
 
     def ui(self, is_img2img):
         original_prompt = gr.Textbox(label="Original prompt", lines=1)
+        original_negative_prompt = gr.Textbox(label="Original negative prompt", lines=1)
         cfg = gr.Slider(label="Decode CFG scale", minimum=0.0, maximum=15.0, step=0.1, value=1.0)
         st = gr.Slider(label="Decode steps", minimum=1, maximum=150, step=1, value=50)
+        randomness = gr.Slider(label="randomness", minimum=0.0, maximum=1.0, step=0.01, value=0.0)
+        return [original_prompt, original_negative_prompt, cfg, st, randomness]
 
-        return [original_prompt, cfg, st]
-
-    def run(self, p, original_prompt, cfg, st):
+    def run(self, p, original_prompt, original_negative_prompt, cfg, st, randomness):
         p.batch_size = 1
         p.batch_count = 1
 
         def sample_extra(x, conditioning, unconditional_conditioning):
             lat = (p.init_latent.cpu().numpy() * 10).astype(int)
 
-            same_params = self.cache is not None and self.cache.cfg_scale == cfg and self.cache.steps == st and self.cache.original_prompt == original_prompt
+            same_params = self.cache is not None and self.cache.cfg_scale == cfg and self.cache.steps == st and self.cache.original_prompt == original_prompt and self.cache.original_negative_prompt == original_negative_prompt
             same_everything = same_params and self.cache.latent.shape == lat.shape and np.abs(self.cache.latent-lat).sum() < 100
 
             if same_everything:
-                noise = self.cache.noise
+                rec_noise = self.cache.noise
             else:
                 shared.state.job_count += 1
                 cond = p.sd_model.get_learned_conditioning(p.batch_size * [original_prompt])
-                uncond = p.sd_model.get_learned_conditioning(p.batch_size * [""])
-                noise = find_noise_for_image(p, cond, uncond, cfg, st)
-                self.cache = Cached(noise, cfg, st, lat, original_prompt)
+                uncond = p.sd_model.get_learned_conditioning(p.batch_size * [original_negative_prompt])
+                rec_noise = find_noise_for_image(p, cond, uncond, cfg, st)
+                self.cache = Cached(rec_noise, cfg, st, lat, original_prompt, original_negative_prompt)
 
+            rand_noise = processing.create_random_tensors(p.init_latent.shape[1:], [p.seed + x + 1 for x in range(p.init_latent.shape[0])])
+            
+            combined_noise = ((1 - randomness) * rec_noise + randomness * rand_noise) / ((randomness**2 + (1-randomness)**2) ** 0.5)
+            
             sampler = samplers[p.sampler_index].constructor(p.sd_model)
 
-            samples_ddim = sampler.sample(p, noise, conditioning, unconditional_conditioning)
-            return samples_ddim
+            sigmas = sampler.model_wrap.get_sigmas(p.steps)
+            
+            noise_dt = combined_noise - ( p.init_latent / sigmas[0] )
+            
+            p.seed = p.seed + 1
+            
+            return sampler.sample_img2img(p, p.init_latent, noise_dt, conditioning, unconditional_conditioning)
+
 
         p.sample = sample_extra
 

scripts/outpainting_mk_2.py

@@ -0,0 +1,290 @@
+import math
+
+import numpy as np
+import skimage
+
+import modules.scripts as scripts
+import gradio as gr
+from PIL import Image, ImageDraw
+
+from modules import images, processing, devices
+from modules.processing import Processed, process_images
+from modules.shared import opts, cmd_opts, state
+
+
+def expand(x, dir, amount, power=0.75):
+    is_left = dir == 3
+    is_right = dir == 1
+    is_up = dir == 0
+    is_down = dir == 2
+
+    if is_left or is_right:
+        noise = np.zeros((x.shape[0], amount, 3), dtype=float)
+        indexes = np.random.random((x.shape[0], amount)) ** power * (1 - np.arange(amount) / amount)
+        if is_right:
+            indexes = 1 - indexes
+        indexes = (indexes * (x.shape[1] - 1)).astype(int)
+
+        for row in range(x.shape[0]):
+            if is_left:
+                noise[row] = x[row][indexes[row]]
+            else:
+                noise[row] = np.flip(x[row][indexes[row]], axis=0)
+
+        x = np.concatenate([noise, x] if is_left else [x, noise], axis=1)
+        return x
+
+    if is_up or is_down:
+        noise = np.zeros((amount, x.shape[1], 3), dtype=float)
+        indexes = np.random.random((x.shape[1], amount)) ** power * (1 - np.arange(amount) / amount)
+        if is_down:
+            indexes = 1 - indexes
+        indexes = (indexes * x.shape[0] - 1).astype(int)
+
+        for row in range(x.shape[1]):
+            if is_up:
+                noise[:, row] = x[:, row][indexes[row]]
+            else:
+                noise[:, row] = np.flip(x[:, row][indexes[row]], axis=0)
+
+        x = np.concatenate([noise, x] if is_up else [x, noise], axis=0)
+        return x
+
+
+def get_matched_noise(_np_src_image, np_mask_rgb, noise_q=1, color_variation=0.05):
+    # helper fft routines that keep ortho normalization and auto-shift before and after fft
+    def _fft2(data):
+        if data.ndim > 2:  # has channels
+            out_fft = np.zeros((data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128)
+            for c in range(data.shape[2]):
+                c_data = data[:, :, c]
+                out_fft[:, :, c] = np.fft.fft2(np.fft.fftshift(c_data), norm="ortho")
+                out_fft[:, :, c] = np.fft.ifftshift(out_fft[:, :, c])
+        else:  # one channel
+            out_fft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
+            out_fft[:, :] = np.fft.fft2(np.fft.fftshift(data), norm="ortho")
+            out_fft[:, :] = np.fft.ifftshift(out_fft[:, :])
+
+        return out_fft
+
+    def _ifft2(data):
+        if data.ndim > 2:  # has channels
+            out_ifft = np.zeros((data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128)
+            for c in range(data.shape[2]):
+                c_data = data[:, :, c]
+                out_ifft[:, :, c] = np.fft.ifft2(np.fft.fftshift(c_data), norm="ortho")
+                out_ifft[:, :, c] = np.fft.ifftshift(out_ifft[:, :, c])
+        else:  # one channel
+            out_ifft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
+            out_ifft[:, :] = np.fft.ifft2(np.fft.fftshift(data), norm="ortho")
+            out_ifft[:, :] = np.fft.ifftshift(out_ifft[:, :])
+
+        return out_ifft
+
+    def _get_gaussian_window(width, height, std=3.14, mode=0):
+        window_scale_x = float(width / min(width, height))
+        window_scale_y = float(height / min(width, height))
+
+        window = np.zeros((width, height))
+        x = (np.arange(width) / width * 2. - 1.) * window_scale_x
+        for y in range(height):
+            fy = (y / height * 2. - 1.) * window_scale_y
+            if mode == 0:
+                window[:, y] = np.exp(-(x ** 2 + fy ** 2) * std)
+            else:
+                window[:, y] = (1 / ((x ** 2 + 1.) * (fy ** 2 + 1.))) ** (std / 3.14)  # hey wait a minute that's not gaussian
+
+        return window
+
+    def _get_masked_window_rgb(np_mask_grey, hardness=1.):
+        np_mask_rgb = np.zeros((np_mask_grey.shape[0], np_mask_grey.shape[1], 3))
+        if hardness != 1.:
+            hardened = np_mask_grey[:] ** hardness
+        else:
+            hardened = np_mask_grey[:]
+        for c in range(3):
+            np_mask_rgb[:, :, c] = hardened[:]
+        return np_mask_rgb
+
+    width = _np_src_image.shape[0]
+    height = _np_src_image.shape[1]
+    num_channels = _np_src_image.shape[2]
+
+    np_src_image = _np_src_image[:] * (1. - np_mask_rgb)
+    np_mask_grey = (np.sum(np_mask_rgb, axis=2) / 3.)
+    img_mask = np_mask_grey > 1e-6
+    ref_mask = np_mask_grey < 1e-3
+
+    windowed_image = _np_src_image * (1. - _get_masked_window_rgb(np_mask_grey))
+    windowed_image /= np.max(windowed_image)
+    windowed_image += np.average(_np_src_image) * np_mask_rgb  # / (1.-np.average(np_mask_rgb))  # rather than leave the masked area black, we get better results from fft by filling the average unmasked color
+
+    src_fft = _fft2(windowed_image)  # get feature statistics from masked src img
+    src_dist = np.absolute(src_fft)
+    src_phase = src_fft / src_dist
+
+    noise_window = _get_gaussian_window(width, height, mode=1)  # start with simple gaussian noise
+    noise_rgb = np.random.random_sample((width, height, num_channels))
+    noise_grey = (np.sum(noise_rgb, axis=2) / 3.)
+    noise_rgb *= color_variation  # the colorfulness of the starting noise is blended to greyscale with a parameter
+    for c in range(num_channels):
+        noise_rgb[:, :, c] += (1. - color_variation) * noise_grey
+
+    noise_fft = _fft2(noise_rgb)
+    for c in range(num_channels):
+        noise_fft[:, :, c] *= noise_window
+    noise_rgb = np.real(_ifft2(noise_fft))
+    shaped_noise_fft = _fft2(noise_rgb)
+    shaped_noise_fft[:, :, :] = np.absolute(shaped_noise_fft[:, :, :]) ** 2 * (src_dist ** noise_q) * src_phase  # perform the actual shaping
+
+    brightness_variation = 0.  # color_variation # todo: temporarily tieing brightness variation to color variation for now
+    contrast_adjusted_np_src = _np_src_image[:] * (brightness_variation + 1.) - brightness_variation * 2.
+
+    # scikit-image is used for histogram matching, very convenient!
+    shaped_noise = np.real(_ifft2(shaped_noise_fft))
+    shaped_noise -= np.min(shaped_noise)
+    shaped_noise /= np.max(shaped_noise)
+    shaped_noise[img_mask, :] = skimage.exposure.match_histograms(shaped_noise[img_mask, :] ** 1., contrast_adjusted_np_src[ref_mask, :], channel_axis=1)
+    shaped_noise = _np_src_image[:] * (1. - np_mask_rgb) + shaped_noise * np_mask_rgb
+
+    matched_noise = shaped_noise[:]
+
+    return np.clip(matched_noise, 0., 1.)
+
+
+
+class Script(scripts.Script):
+    def title(self):
+        return "Outpainting mk2"
+
+    def show(self, is_img2img):
+        return is_img2img
+
+    def ui(self, is_img2img):
+        if not is_img2img:
+            return None
+
+        info = gr.HTML("<p style=\"margin-bottom:0.75em\">Recommended settings: Sampling Steps: 80-100, Sampler: Euler a, Denoising strength: 0.8</p>")
+
+        pixels = gr.Slider(label="Pixels to expand", minimum=8, maximum=256, step=8, value=128)
+        mask_blur = gr.Slider(label='Mask blur', minimum=0, maximum=64, step=1, value=8, visible=False)
+        direction = gr.CheckboxGroup(label="Outpainting direction", choices=['left', 'right', 'up', 'down'], value=['left', 'right', 'up', 'down'])
+        noise_q = gr.Slider(label="Fall-off exponent (lower=higher detail)", minimum=0.0, maximum=4.0, step=0.01, value=1.0)
+        color_variation = gr.Slider(label="Color variation", minimum=0.0, maximum=1.0, step=0.01, value=0.05)
+
+        return [info, pixels, mask_blur, direction, noise_q, color_variation]
+
+    def run(self, p, _, pixels, mask_blur, direction, noise_q, color_variation):
+        initial_seed_and_info = [None, None]
+
+        process_width = p.width
+        process_height = p.height
+
+        p.mask_blur = mask_blur*4
+        p.inpaint_full_res = False
+        p.inpainting_fill = 1
+        p.do_not_save_samples = True
+        p.do_not_save_grid = True
+
+        left = pixels if "left" in direction else 0
+        right = pixels if "right" in direction else 0
+        up = pixels if "up" in direction else 0
+        down = pixels if "down" in direction else 0
+
+        init_img = p.init_images[0]
+        target_w = math.ceil((init_img.width + left + right) / 64) * 64
+        target_h = math.ceil((init_img.height + up + down) / 64) * 64
+
+        if left > 0:
+            left = left * (target_w - init_img.width) // (left + right)
+        if right > 0:
+            right = target_w - init_img.width - left
+
+        if up > 0:
+            up = up * (target_h - init_img.height) // (up + down)
+
+        if down > 0:
+            down = target_h - init_img.height - up
+
+        init_image = p.init_images[0]
+
+        state.job_count = (1 if left > 0 else 0) + (1 if right > 0 else 0)+ (1 if up > 0 else 0)+ (1 if down > 0 else 0)
+
+        def expand(init, expand_pixels, is_left=False, is_right=False, is_top=False, is_bottom=False):
+            is_horiz = is_left or is_right
+            is_vert = is_top or is_bottom
+            pixels_horiz = expand_pixels if is_horiz else 0
+            pixels_vert = expand_pixels if is_vert else 0
+
+            img = Image.new("RGB", (init.width + pixels_horiz, init.height + pixels_vert))
+            img.paste(init, (pixels_horiz if is_left else 0, pixels_vert if is_top else 0))
+            mask = Image.new("RGB", (init.width + pixels_horiz, init.height + pixels_vert), "white")
+            draw = ImageDraw.Draw(mask)
+            draw.rectangle((
+                expand_pixels + mask_blur if is_left else 0,
+                expand_pixels + mask_blur if is_top else 0,
+                mask.width - expand_pixels - mask_blur if is_right else mask.width,
+                mask.height - expand_pixels - mask_blur if is_bottom else mask.height,
+            ), fill="black")
+
+            np_image = (np.asarray(img) / 255.0).astype(np.float64)
+            np_mask = (np.asarray(mask) / 255.0).astype(np.float64)
+            noised = get_matched_noise(np_image, np_mask, noise_q, color_variation)
+            out = Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB")
+
+            target_width = min(process_width, init.width + pixels_horiz) if is_horiz else img.width
+            target_height = min(process_height, init.height + pixels_vert) if is_vert else img.height
+
+            crop_region = (
+                0 if is_left else out.width - target_width,
+                0 if is_top else out.height - target_height,
+                target_width if is_left else out.width,
+                target_height if is_top else out.height,
+            )
+
+            image_to_process = out.crop(crop_region)
+            mask = mask.crop(crop_region)
+
+            p.width = target_width if is_horiz else img.width
+            p.height = target_height if is_vert else img.height
+            p.init_images = [image_to_process]
+            p.image_mask = mask
+
+            latent_mask = Image.new("RGB", (p.width, p.height), "white")
+            draw = ImageDraw.Draw(latent_mask)
+            draw.rectangle((
+                expand_pixels + mask_blur * 2 if is_left else 0,
+                expand_pixels + mask_blur * 2 if is_top else 0,
+                mask.width - expand_pixels - mask_blur * 2 if is_right else mask.width,
+                mask.height - expand_pixels - mask_blur * 2 if is_bottom else mask.height,
+            ), fill="black")
+            p.latent_mask = latent_mask
+
+            proc = process_images(p)
+            proc_img = proc.images[0]
+
+            if initial_seed_and_info[0] is None:
+                initial_seed_and_info[0] = proc.seed
+                initial_seed_and_info[1] = proc.info
+
+            out.paste(proc_img, (0 if is_left else out.width - proc_img.width, 0 if is_top else out.height - proc_img.height))
+            return out
+
+        img = init_image
+
+        if left > 0:
+            img = expand(img, left, is_left=True)
+        if right > 0:
+            img = expand(img, right, is_right=True)
+        if up > 0:
+            img = expand(img, up, is_top=True)
+        if down > 0:
+            img = expand(img, down, is_bottom=True)
+
+        res = Processed(p, [img], initial_seed_and_info[0], initial_seed_and_info[1])
+
+        if opts.samples_save:
+            images.save_image(img, p.outpath_samples, "", res.seed, p.prompt, opts.grid_format, info=res.info, p=p)
+
+        return res
+

scripts/prompts_from_file.py

@@ -13,28 +13,42 @@ from modules.shared import opts, cmd_opts, state
 
 class Script(scripts.Script):
     def title(self):
-        return "Prompts from file"
+        return "Prompts from file or textbox"
 
     def ui(self, is_img2img):
+        # This checkbox would look nicer as two tabs, but there are two problems:
+        # 1) There is a bug in Gradio 3.3 that prevents visibility from working on Tabs
+        # 2) Even with Gradio 3.3.1, returning a control (like Tabs) that can't be used as input
+        #    causes a AttributeError: 'Tabs' object has no attribute 'preprocess' assert,
+        #    due to the way Script assumes all controls returned can be used as inputs.
+        # Therefore, there's no good way to use grouping components right now,
+        # so we will use a checkbox! :)
+        checkbox_txt = gr.Checkbox(label="Show Textbox", value=False)
         file = gr.File(label="File with inputs", type='bytes')
+        prompt_txt = gr.TextArea(label="Prompts")
+        checkbox_txt.change(fn=lambda x: [gr.File.update(visible = not x), gr.TextArea.update(visible = x)], inputs=[checkbox_txt], outputs=[file, prompt_txt])
+        return [checkbox_txt, file, prompt_txt]
 
-        return [file]
-
-    def run(self, p, data: bytes):
-        lines = [x.strip() for x in data.decode('utf8', errors='ignore').split("\n")]
+    def run(self, p, checkbox_txt, data: bytes, prompt_txt: str):
+        if (checkbox_txt):
+            lines = [x.strip() for x in prompt_txt.splitlines()]
+        else:
+            lines = [x.strip() for x in data.decode('utf8', errors='ignore').split("\n")]
         lines = [x for x in lines if len(x) > 0]
 
-        batch_count = math.ceil(len(lines) / p.batch_size)
-        print(f"Will process {len(lines) * p.n_iter} images in {batch_count * p.n_iter} batches.")
+        img_count = len(lines) * p.n_iter
+        batch_count = math.ceil(img_count / p.batch_size)
+        loop_count = math.ceil(batch_count / p.n_iter)
+        print(f"Will process {img_count} images in {batch_count} batches.")
 
         p.do_not_save_grid = True
 
         state.job_count = batch_count
 
         images = []
-        for batch_no in range(batch_count):
-            state.job = f"{batch_no + 1} out of {batch_count * p.n_iter}"
-            p.prompt = lines[batch_no*p.batch_size:(batch_no+1)*p.batch_size] * p.n_iter
+        for loop_no in range(loop_count):
+            state.job = f"{loop_no + 1} out of {loop_count}"
+            p.prompt = lines[loop_no*p.batch_size:(loop_no+1)*p.batch_size] * p.n_iter
             proc = process_images(p)
             images += proc.images
 

scripts/xy_grid.py

@@ -10,7 +10,9 @@ import gradio as gr
 from modules import images
 from modules.processing import process_images, Processed
 from modules.shared import opts, cmd_opts, state
+import modules.shared as shared
 import modules.sd_samplers
+import modules.sd_models
 import re
 
 
@@ -41,6 +43,15 @@ def apply_sampler(p, x, xs):
     p.sampler_index = sampler_index
 
 
+def apply_checkpoint(p, x, xs):
+    applicable = [info for info in modules.sd_models.checkpoints_list.values() if x in info.title]
+    assert len(applicable) > 0, f'Checkpoint {x} for found'
+
+    info = applicable[0]
+
+    modules.sd_models.reload_model_weights(shared.sd_model, info)
+
+
 def format_value_add_label(p, opt, x):
     if type(x) == float:
         x = round(x, 8)
@@ -74,15 +85,16 @@ axis_options = [
     AxisOption("CFG Scale", float, apply_field("cfg_scale"), format_value_add_label),
     AxisOption("Prompt S/R", str, apply_prompt, format_value),
     AxisOption("Sampler", str, apply_sampler, format_value),
+    AxisOption("Checkpoint name", str, apply_checkpoint, format_value),
     AxisOptionImg2Img("Denoising", float, apply_field("denoising_strength"), format_value_add_label), #  as it is now all AxisOptionImg2Img items must go after AxisOption ones
 ]
 
 
-def draw_xy_grid(p, xs, ys, x_label, y_label, cell, draw_legend):
+def draw_xy_grid(p, xs, ys, x_labels, y_labels, cell, draw_legend):
     res = []
 
-    ver_texts = [[images.GridAnnotation(y_label(y))] for y in ys]
-    hor_texts = [[images.GridAnnotation(x_label(x))] for x in xs]
+    ver_texts = [[images.GridAnnotation(y)] for y in y_labels]
+    hor_texts = [[images.GridAnnotation(x)] for x in x_labels]
 
     first_pocessed = None
 
@@ -206,8 +218,8 @@ class Script(scripts.Script):
             p,
             xs=xs,
             ys=ys,
-            x_label=lambda x: x_opt.format_value(p, x_opt, x),
-            y_label=lambda y: y_opt.format_value(p, y_opt, y),
+            x_labels=[x_opt.format_value(p, x_opt, x) for x in xs],
+            y_labels=[y_opt.format_value(p, y_opt, y) for y in ys],
             cell=cell,
             draw_legend=draw_legend
         )
@@ -215,4 +227,7 @@ class Script(scripts.Script):
         if opts.grid_save:
             images.save_image(processed.images[0], p.outpath_grids, "xy_grid", prompt=p.prompt, seed=processed.seed, grid=True, p=p)
 
+        # restore checkpoint in case it was changed by axes
+        modules.sd_models.reload_model_weights(shared.sd_model)
+
         return processed