--- title: "Generating Full-Length Songs with YuE2 BF16: Overcoming Duration Limits and Audio Truncation" description: "AI music tools cut off complex tracks mid-sentence. Here is how I solved the 240s duration wall using YuE2 BF16, 295s headroom budgeting, and a 60-song pipeline." pubDate: 2026-10-26 category: tutorials blueprint: ai-agent-automation blueprintLayer: L4-aigc-media-pipelines stream: creative-lab tags: ["ai-music", "yue2", "comfyui", "audio-engineering", "subtitles", "python"] ogImage: "/og/generating-full-length-songs-with-yue2-comfyui.jpg" banner: "/banners/generating-full-length-songs-with-yue2-comfyui.jpg" draft: false dynamicCta: true --- Every generative music engine I have benchmarked suffers from the same fatal flaw: **the sudden, unceremonious cutoff**. Whether you use cloud APIs or local weights, models trained on fixed acoustic token windows struggle the moment you give them a complete, narrative song structure. A 9-part progression—`[Intro] -> [Verse 1] -> [Chorus] -> [Interlude] -> [Verse 2] -> [Chorus] -> [Bridge] -> [Chorus] -> [Outro]`—inevitably collides with the hard ceiling of the generation buffer. The vocalist is mid-word in the climax, and the audio abruptly flatlines into digital silence. After iterating across hundreds of takes on a dedicated 12GB VRAM on-premise inference node (RTX 3060), I architected an end-to-end production pipeline that stably generates full-length songs spanning 3 to 4.7 minutes (up to 283.1s) with complete 9-stage progression, natural acoustic fade-outs, and zero mid-phrase amputations. Below is the breakdown of why standard generation fails, the 295-second headroom formula that fixes it, and the full collection of **60 completed full-length songs** (with individual 1024×1024 cover art and synchronized scrolling lyrics). ---
## The Problem: The 240-Second Wall In earlier batches, songs generated via standard ComfyUI text-to-music nodes frequently exhibited one of two failure modes: 1. **The Chopped Outro**: Setting `max_duration = 240.0s` while feeding a full lyric sheet resulted in the track hitting the final timestamp while the acoustic model was still executing the chorus. All 12 initial test runs ended at exactly `240.000s`—an unusable hard cliff. 2. **The Panic Accelerando**: If you prompt the model to fit too many lines into a tight window, the auto-regressive acoustic token predictor compresses the vocal pacing, turning an emotional Celtic ballad into an unnatural rap cadence. The fundamental mistake was treating `max_duration` as the *desired song length*. In generative auto-regressive audio, **max duration is not your target length; it is your acoustic safety wall.** ## The Three Architecture Fixes | Engineering Metric | V1 Naive Pipeline (`max_duration=240s`) | V2 Production Pipeline (`295s Headroom + Gate`) | Architecture Gain | | :--- | :--- | :--- | :--- | | **Natural Fade-Out Yield** | 0 / 12 (0.0%) | 60 / 60 (100.0% Gate Passed) | Eliminates manual acoustic auditing costs | | **Finished Duration Distribution** | Deadlocked at `240.000s` | `129.9s ~ 283.1s` (up to 4m43s) | Breathes naturally with genre BPM and dynamics | | **Continuous Batch VRAM Stability** | CUDA OOM crash at track 4~5 | 60 tracks unattended with zero failures | Proactive VRAM defragmentation lifecycle | ### 1. The 295.0s Headroom Formula To produce songs that naturally conclude around a 4-minute runtime, you must structure the lyric density for **210s to 260s**, while provisioning the generator with a **295.0s** ceiling: $$ ext{Headroom Buffer} = ext{Max Duration (295s)} - ext{Expected Lyric Duration (220s)} pprox 75 ext{s}$$ This ~75-second buffer provides the acoustic model with the token budget required to: - Complete the final verse without panic compression; - Introduce a dedicated, melodic instrumental outro; - Execute natural acoustic decay, room reverb dissipation, and final silence. ### 2. Automated Hard-Cut Quarantine Gate In an unattended batch pipeline, human listening checks do not scale. Any take whose generated duration measures within 1.0s of the hard ceiling (`>= 294.0s`) or exhibits missing end-of-sequence silence is treated as an unresolved truncation failure and automatically quarantined for re-queueing with a new seed. Only takes that resolve naturally before the ceiling (`120.0s <= duration < 294.0s`) pass into the mastering and LRC alignment stage. Across our 60-song production run, all 60 finalized masters resolved cleanly between **129.9s and 283.1s** with zero abrupt cutoffs.
pipeline/quarantine_gate.py (38 lines) — Automated ffprobe duration gate & ComfyUI VRAM flush ```python import json import subprocess import urllib.request from pathlib import Path CEILING_SECONDS = 295.0 QUARANTINE_THRESHOLD = 294.0 MIN_VALID_DURATION = 120.0 COMFY_API_URL = "http://127.0.0.1:8188" def probe_audio_duration(audio_path: Path) -> float: cmd = [ "ffprobe", "-v", "error", "-show_entries", "format=duration", "-of", "json", str(audio_path) ] result = subprocess.run(cmd, capture_output=True, text=True, check=True) payload = json.loads(result.stdout) return float(payload["format"]["duration"]) def release_comfy_vram() -> None: req = urllib.request.Request( f"{COMFY_API_URL}/free", data=json.dumps({"unload_models": True, "free_memory": True}).encode("utf-8"), headers={"Content-Type": "application/json"}, method="POST", ) urllib.request.urlopen(req, timeout=10) def validate_and_gate_take(audio_path: Path, quarantine_dir: Path) -> bool: duration = probe_audio_duration(audio_path) release_comfy_vram() if duration >= QUARANTINE_THRESHOLD or duration < MIN_VALID_DURATION: target = quarantine_dir / f"REJECTED_{duration:.1f}s_{audio_path.name}" audio_path.rename(target) return False return True ```
### 3. VRAM Lifecycle Management on a 12GB Production Node Running YuE2-3B in BF16 on a cost-optimized 12GB VRAM self-hosted production node requires strict memory lifecycle hygiene to prevent fragmentation across multi-hour runs: ```bash # Launch ComfyUI production worker with deterministic VRAM allocation cd /opt/ai-pipelines/comfyui-worker ./python_embeded/python.exe -s ComfyUI/main.py \ --windows-standalone-build \ --port 8188 \ --disable-pinned-memory \ --cache-classic ``` By enforcing immediate model unloading via the `/free` endpoint after each batch execution and eliminating pinned memory overhead, continuous generation of 4.5-minute sequences remains rock-solid without triggering CUDA out-of-memory faults. --- ## The 60 Full-Length Song Archive All 60 tracks below were generated locally with YuE2 BF16, tagged with clean ID3v2.3 metadata, paired with custom 1024×1024 Gemini-rendered full-bleed cover art (converted to WebP), and hosted on CDN with HTTP 206 range-seeking support. *Tip: Use the interactive selector above to jump directly to any track.*
The Thimble Under the Moss cover art
01. The Thimble Under the Moss MP3 3m 11s (191.1s) · Celtic Folk · Crisp Youthful Treble Vocal (Cécile Corbel style) | Upbeat Tempo (~105 BPM)
Mist of Skellig Michael cover art
02. Mist of Skellig Michael MP3 4m 26s (266.2s) · Celtic Folk · Ethereal Youthful Solo (Enya style with pure treble vocal) | Slow & Atmospheric (~72 BPM)
The Rose of Connemara cover art
03. The Rose of Connemara MP3 4m 43s (283.1s) · Celtic Folk · Pure Youthful Solo (Early Cara Dillon style) | Mid-Slow Narrative (~82 BPM)
Lanterns on the Moor cover art
04. Lanterns on the Moor MP3 3m 31s (211.2s) · Celtic Folk · Playful Sprite-like Female Vocal (Upbeat Cécile Corbel style) | Fast 3/4 Waltz (~128 BPM)
The Weaver of Raglan Lane cover art
05. The Weaver of Raglan Lane MP3 4m 39s (279.3s) · Celtic Folk · Warm Resonant Celtic Folk Baritone (Luke Kelly style) | Mid-Slow (~80 BPM)
硝子の花束 cover art
06. 硝子の花束 MP3 3m 49s (230.0s) · Japanese Healing · Pure Airy Anime Soprano with Delicate Vulnerability (Inori style) | Mid-Slow Piano & Strings (~78 BPM)
銀の時計塔と宵の歌 cover art
07. 銀の時計塔と宵の歌 MP3 4m 34s (274.8s) · Japanese Healing · Luminous Soprano Vocal with Kajiura-style Chanting | Mid-Tempo 3/4 Waltz (~110 BPM)
森の郵便屋さん cover art
08. 森の郵便屋さん MP3 3m 44s (225.0s) · Japanese Healing · Ultra-Gentle Whispered Youthful Vocal | Slow Acoustic Guitar & Harp (~68 BPM)
白鳥の祈り cover art
09. 白鳥の祈り MP3 4m 06s (246.3s) · Japanese Healing · Ethereal High Soprano with Expansive Reverb | Slow Cinematic Orchestral (~70 BPM)
星屑のオルゴール cover art
10. 星屑のオルゴール MP3 3m 31s (211.8s) · Japanese Healing · Sweet Youthful Anime Soprano | Mid-Slow Light Ballad (~88 BPM)
ひまわり行進曲 cover art
11. ひまわり行進曲 MP3 3m 18s (198.0s) · Japanese Energetic · High-Energy Youthful Solo with Chorus Accents | Fast Upbeat Pop (~118 BPM)
プリンは世界を救うのだ! cover art
12. プリンは世界を救うのだ! MP3 3m 23s (203.4s) · Japanese Energetic · Bright Anime-Style Soprano Vocal (Fast, Dense Cadence) | High-Tempo Electro-Pop (~145 BPM)
こねこ探偵団のテーマ cover art
13. こねこ探偵団のテーマ MP3 2m 16s (136.1s) · Japanese Energetic · Playful Childlike Vocal (Ages 6–8 Character Voice) | Upbeat Swing Nursery Rhyme (~115 BPM)
夏祭りのラムネ星 cover art
14. 夏祭りのラムネ星 MP3 3m 14s (195.0s) · Japanese Energetic · Bright Youthful Duet (Tenor & Soprano Harmonies) | Fast Summer Anime Pop (~122 BPM)
ほうきに乗って宅配便 cover art
15. ほうきに乗って宅配便 MP3 3m 16s (196.1s) · Japanese Energetic · Crisp Sunny Youthful Vocal | Mid-Tempo Vintage J-Pop (~116 BPM)
白露·竹间灯 cover art
16. 白露·竹间灯 MP3 3m 42s (222.6s) · Chinese Guofeng · Sweet Melodic Youthful Vocal | Mid-Tempo Electro-Guofeng (~108 BPM)
小鲤跃龙门 cover art
17. 小鲤跃龙门 MP3 2m 09s (129.9s) · Chinese Guofeng · Lively Playful Treble Vocal (Fast Lyric Cadence) | Mid-Fast Folk Pop (~120 BPM)
云海归鲲 cover art
18. 云海归鲲 MP3 3m 56s (236.9s) · Chinese Guofeng · Ethereal Classical Falsetto / Bel Canto Male Vocal | Slow Ambient Ballad (~68 BPM)
纸伞巷 cover art
19. 纸伞巷 MP3 3m 34s (214.6s) · Chinese Guofeng · Gentle Melodic Soprano with Jiangnan Ditty Charm | Mid-Slow Guofeng (~80 BPM)
山间小道士 cover art
20. 山间小道士 MP3 3m 07s (187.6s) · Chinese Guofeng · Pure Unadorned Youthful Vocal | Mid-Tempo Guofeng Acoustic (~105 BPM)
橘子汽水的夏天 cover art
21. 橘子汽水的夏天 MP3 2m 48s (169.0s) · Gentle Acoustic · Clear Youthful Solo Vocal | Mid-Slow Acoustic Guitar Ballad (~82 BPM)
外婆的小院 cover art
22. 外婆的小院 MP3 3m 32s (212.4s) · Gentle Acoustic · Tender Gentle Childlike Vocal | Slow Nostalgic Ballad (~72 BPM)
开往北方的绿皮车 cover art
23. 开往北方的绿皮车 MP3 3m 36s (216.3s) · Gentle Acoustic · Deep Warm Folk Baritone | Slow Acoustic Narrative (~75 BPM)
写给流浪猫的信 cover art
24. 写给流浪猫的信 MP3 2m 09s (129.9s) · Gentle Acoustic · Playful Tender Youthful Vocal | Upbeat Ukulele & Acoustic Guitar (~98 BPM)
平凡的傍晚 cover art
25. 平凡的傍晚 MP3 3m 41s (221.0s) · Gentle Acoustic · Gentle Healing Young Male Vocal | Slow Piano & Acoustic Guitar (~74 BPM)
Light of the Eventide cover art
26. Light of the Eventide MP3 4m 09s (249.0s) · Angelic Choir · Sublime Treble Solo with Pure Youth Choir | Solemn Hymn (~66 BPM)
Bells of the Winter Star cover art
27. Bells of the Winter Star MP3 2m 47s (167.6s) · Angelic Choir · Crisp Multi-Part Youth Choir Polyphony & Canon | Dynamic 3/4 Meter (~125 BPM)
Paper Planes in the Wind cover art
28. Paper Planes in the Wind MP3 3m 25s (205.5s) · Angelic Choir · Luminous Youth Solo with Collegiate Choir Harmonies | Warm Mid-Tempo (~96 BPM)
Lullaby of the Little Star cover art
29. Lullaby of the Little Star MP3 3m 47s (227.8s) · Angelic Choir · Pure Unadorned Youthful Solo | Gentle Lullaby (~70 BPM)
Hymn of the Morning Meadow cover art
30. Hymn of the Morning Meadow MP3 4m 19s (259.9s) · Angelic Choir · Sacred Treble Solo with Expansive Choral Harmony (Pachelbel Canon Progression) | Slow Uplifting (~72 BPM)
地板下的借物小人 cover art
31. 地板下的借物小人 MP3 3m 07s (187.7s) · Celtic Folk (ZH) · Crisp Youthful Treble Vocal | Upbeat Celtic Harp & Tin Whistle (~105 BPM)
迷雾孤岛的竖琴 cover art
32. 迷雾孤岛的竖琴 MP3 3m 24s (204.4s) · Celtic Folk (ZH) · Ethereal Youthful Solo | Slow & Atmospheric Celtic Ballad (~70 BPM)
石楠花开的高地 cover art
33. 石楠花开的高地 MP3 3m 36s (216.9s) · Celtic Folk (ZH) · Pure Youthful Solo | Mid-Slow Acoustic Guitar & Celtic Fiddle (~82 BPM)
橡树下的萤火舞会 cover art
34. 橡树下的萤火舞会 MP3 3m 43s (223.1s) · Celtic Folk (ZH) · Playful Sprite-like Female Vocal (Fast Cadence) | Fast 3/4 Waltz (~128 BPM)
老航海家的绿宝石 cover art
35. 老航海家的绿宝石 MP3 3m 49s (229.8s) · Celtic Folk (ZH) · Warm Resonant Celtic Folk Baritone | Mid-Slow Acoustic (~80 BPM)
玻璃温室的告别 cover art
36. 玻璃温室的告别 MP3 3m 37s (217.0s) · Japanese Healing (ZH) · Pure Airy Anime Soprano with Delicate Vulnerability | Mid-Slow Piano & Strings (~78 BPM)
银色钟塔与暮光咏叹 cover art
37. 银色钟塔与暮光咏叹 MP3 4m 01s (241.4s) · Japanese Healing (ZH) · Luminous Soprano Vocal with Conlang Chants | Classical Gothic Waltz (~110 BPM)
森林邮差的蒲公英 cover art
38. 森林邮差的蒲公英 MP3 3m 26s (206.7s) · Japanese Healing (ZH) · Ultra-Gentle Whispered Youthful Vocal | Slow Acoustic Guitar & Celesta (~68 BPM)
白鸟与雪原的祈愿 cover art
39. 白鸟与雪原的祈愿 MP3 3m 39s (220.0s) · Japanese Healing (ZH) · Ethereal High Soprano with Expansive Reverb | Slow Cinematic Orchestral (~70 BPM)
发条星尘八音盒 cover art
40. 发条星尘八音盒 MP3 3m 18s (198.2s) · Japanese Healing (ZH) · Sweet Youthful Anime Soprano | Mid-Slow Music Box & Guitar (~88 BPM)
向日葵小队出发 cover art
41. 向日葵小队出发 MP3 2m 39s (159.4s) · Japanese Energetic (ZH) · High-Energy Youth Solo with Chorus Accents | Fast Upbeat Pop (~118 BPM)
焦糖布丁保卫战! cover art
42. 焦糖布丁保卫战! MP3 3m 38s (218.9s) · Japanese Energetic (ZH) · Bright Anime-Style Soprano Vocal (Fast, Dense Cadence) | High-Tempo Electro-Pop (~145 BPM)
猫咪侦探事务所 cover art
43. 猫咪侦探事务所 MP3 2m 31s (151.4s) · Japanese Energetic (ZH) · Playful Childlike Vocal (Ages 6–8 Character Voice) | Upbeat Xylophone & Brass (~115 BPM)
夏夜烟火与波子汽水 cover art
44. 夏夜烟火与波子汽水 MP3 3m 07s (187.8s) · Japanese Energetic (ZH) · Bright Youthful Duet (Tenor & Soprano Harmonies) | Fast Summer Anime Pop (~122 BPM)
骑上扫帚送快递 cover art
45. 骑上扫帚送快递 MP3 3m 08s (188.5s) · Japanese Energetic (ZH) · Crisp Sunny Youthful Vocal | Mid-Tempo Guitar & Tambourine (~116 BPM)
小满·青梅煮雨 cover art
46. 小满·青梅煮雨 MP3 3m 34s (214.9s) · Chinese Guofeng New · Sweet Melodic Youthful Vocal | Mid-Tempo Electro-Guofeng (~106 BPM)
糖画摊前的小狐仙 cover art
47. 糖画摊前的小狐仙 MP3 3m 12s (192.2s) · Chinese Guofeng New · Lively Playful Treble Vocal (Fast Lyric Cadence) | Mid-Fast Guofeng (~124 BPM)
鲛人泪·沧海月 cover art
48. 鲛人泪·沧海月 MP3 3m 59s (239.8s) · Chinese Guofeng New · Ethereal Classical Falsetto / Bel Canto Male Vocal | Slow Ambient Guofeng (~68 BPM)
兰亭旧墨 cover art
49. 兰亭旧墨 MP3 3m 49s (229.3s) · Chinese Guofeng New · Gentle Melodic Soprano | Mid-Slow Guzheng & Erhu (~80 BPM)
采药小童下山记 cover art
50. 采药小童下山记 MP3 3m 24s (204.4s) · Chinese Guofeng New · Pure Unadorned Youthful Vocal | Mid-Tempo Bamboo Flute & Guitar (~104 BPM)
晒太阳的旧毛衣 cover art
51. 晒太阳的旧毛衣 MP3 3m 04s (184.9s) · Gentle Acoustic New · Clear Youthful Solo Vocal | Mid-Slow Acoustic Guitar Ballad (~80 BPM)
奶奶的桂花糕 cover art
52. 奶奶的桂花糕 MP3 3m 42s (222.4s) · Gentle Acoustic New · Tender Gentle Childlike Vocal | Slow Harmonica & Acoustic Guitar (~72 BPM)
北方的末班渡轮 cover art
53. 北方的末班渡轮 MP3 3m 49s (229.6s) · Gentle Acoustic New · Deep Warm Folk Baritone | Slow Acoustic Narrative (~74 BPM)
云朵收集指南 cover art
54. 云朵收集指南 MP3 2m 56s (176.4s) · Gentle Acoustic New · Playful Tender Youthful Vocal | Mid-Tempo Gentle Acoustic (~98 BPM)
街角二十四小时书店 cover art
55. 街角二十四小时书店 MP3 4m 13s (253.4s) · Gentle Acoustic New · Gentle Healing Young Male Vocal | Slow Piano & Acoustic Guitar (~74 BPM)
暮光归途 cover art
56. 暮光归途 MP3 4m 02s (242.6s) · Angelic Choir (ZH) · Sublime Treble Solo with Pure Youth Choir | Solemn Pipe Organ & Strings (~66 BPM)
冬夜银铃 cover art
57. 冬夜银铃 MP3 2m 59s (179.0s) · Angelic Choir (ZH) · Crisp Multi-Part Youth Choir Polyphony & Canon | Dynamic 3/4 Meter (~125 BPM)
迎风飞翔的纸飞机 cover art
58. 迎风飞翔的纸飞机 MP3 3m 53s (233.2s) · Angelic Choir (ZH) · Luminous Youth Solo with Collegiate Choir Harmonies | Warm Mid-Tempo (~96 BPM)
小星星的晚安曲 cover art
59. 小星星的晚安曲 MP3 3m 51s (232.0s) · Angelic Choir (ZH) · Pure Unadorned Youthful Solo | Slow Harp & Violin Lullaby (~70 BPM)
晨曦草地的赞美诗 cover art
60. 晨曦草地的赞美诗 MP3 4m 26s (266.0s) · Angelic Choir (ZH) · Sacred Treble Solo with Expansive Choral Harmony (Canon Progression) | Slow Uplifting (~72 BPM)
--- ## Need an Automated AI Media Pipeline for Your Business? Cloud generative APIs charge per-second or per-generation fees that quickly erode margins at scale—and still fail when you need deterministic duration control, batch quality gates, or strict brand consistency. I architect and deploy **self-hosted, zero-marginal-cost AI media and automation pipelines** on dedicated on-premise or private cloud infrastructure for SMEs and digital product teams: - **Unattended AIGC Media Factories**: Automated audio, voiceover, subtitle alignment, and video transcoding pipelines with programmatic quality quarantine gates. - **Predictable Unit Economics**: Run high-throughput BF16/FP8 inference on cost-optimized 12GB–24GB GPU nodes without recurring third-party API lock-in. - **Full-Stack Delivery**: From backend job queues and Dockerized microservices to CDN range-streaming and interactive web storefronts. **Let's discuss your infrastructure or automation bottleneck:** - **WhatsApp**: [+60 12-797 2969](https://wa.me/60127972969) - **Email**: [me@hoelee.com](mailto:me@hoelee.com?subject=Inquiry%20regarding%20AI%20Media%20Pipeline%20Architecture) - **Portfolio & Systems**: [hoelee.com](https://hoelee.com)