# 10-min animated episodes made 100% locally: Qwen-Image + Qwen-Edit face sheets → Wan 2.x I2V (4-step Lightning) → InfiniteTalk lip-sync → Indic Parler-TTS. Workflow + lessons inside — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/a82d25aa838ce390

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-26 07:30 ingested 2026-10-27 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xfq81y/10min_animated_episodes_made_100_locally/)

## retained text



```
This is a clip from Episode 3 (\~10 min) of a 2.5D animated cartoon biography of the Indian sage Valmiki: a hunting trip that goes wrong when a leopard attacks at night. The main version is Hindi, and the episode also has Gujarati, English and other Indian-language audio tracks. This clip uses the English dub, so the lip-sync (driven by the Hindi track) won't match the English words exactly. Every step uses open-weight models running locally on one AMD MI300X (ROCm, ComfyUI), with an M1 Max Mac for upscaling and audio. No paid APIs anywhere.

**Workflow**

1. **Character face sheets:** Qwen-Image, text-to-image on a plain background, one sheet per age stage. Each character has a fixed identity line (skin, face shape, eyes, brows, nose, one distinct mark) repeated verbatim in every prompt. Chaining Qwen-Edit from one age sheet to the next drifted and added noise at each hop, so each stage is generated fresh.
2. **Shot stills:** Qwen-Image (Lightning LoRA, 8 steps) for scenery. Qwen-Image-Edit + face sheet + identity line for named characters. Group wide shots are text-only, with each extra's position spelled out and "different faces" in the prompt; otherwise the face reference bleeds into the extras and you get clones.
3. **Motion:** Wan 2.x I2V at native 720p with the Lightning step-distill LoRA (4 steps). One render per shot, about 4 min for a 7 s shot instead of 45-60. I gate every clip on sharpness against its still and re-roll once if it's soft.
4. **Lip-sync:** InfiniteTalk (4 steps) for Episodes 1-5, close-ups only, driven by the Hindi track (this clip). From Episode 6, talking shots use **LongCat-Video-Avatar 1.5** at 720p: clearer mouth shapes and real head and body movement, but about 4-9x slower (\~1.5 min of GPU per second of speech). On AMD it needs the slower standard attention path, because the fast one produced noise.
5. **Voices:** AI4Bharat Indic Parler-TTS (Mini; Large gave noise on ROCm). Each line is levelled to -16 LUFS before mixing.
6. **Music/SFX:** ACE-Step 1.5 for the background score, ducked 10-12 dB under speech.
7. **Finishing:** Real-ESRGAN animevideov3 x2 on the Mac, plus Blender/depth parallax for shots with no people. Assembly, the subtitle line-check and the mix are Python + ffmpeg.

**Lessons (the expensive ones)**

* **Negations in the positive prompt get drawn.** "No people, no poles" put people and poles in 6 of 7 renders. Exclusions go only in the negative prompt.
* **InfiniteTalk animates every visible face,** including listeners and babies, even in multi-person mode with a silent track. Pasting closed-mouth faces back in caused doubled faces. What worked: a calm wide shot with no lip-sync, then a separate close-up still of only the speaker for the lip-synced line.
* **480p + upscale distorted the eyes.** Native 720p and then ESRGAN x2 kept faces sharp.
* **Don't crop more than \~1.5x from a still.** Tight crops from wides looked soft and sometimes put a line on the wrong character. Close-ups need their own close-up stills.
* **Culture-specific prompts need care.** "Ancient city" gave Mughal domes and European spires, so I prompt Nagara shikhara temple towers and put dome/minaret/spire in the negative. "Bullock cart" gave an American covered wagon until I spelled it out: two white humped bullocks, solid wooden wheels, open flat bed.
* **Two ComfyUI servers on one GPU** gained only \~10-20%. The real limit was pod RAM (each instance holds \~40 GB of model copies), and mixing model families in parallel caused OOMs.

Throughput is now about 3.5-4 GPU-hours of picture per episode.

Happy to share settings or answer questions. Critique on motion and lip-sync is very welcome.
```



document fetched 2026-10-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xfq81y/10min_animated_episodes_made_100_locally
