Fully local open-weight pipeline produces 10-min animated episodes at ~4 GPU-hours each

Zangetsu__0 · reddit · 2026-10-26

A creator produced 10-minute 2.5D animated episodes (a biography of the Indian sage Valmiki, with Hindi/Gujarati/English tracks) entirely with open-weight models on one AMD MI300X via ROCm + ComfyUI — no paid APIs.

Workflow

Qwen-Image character face sheets per age stage with a verbatim identity line; fresh generation per stage (Qwen-Edit chains drift)
Shot stills: Qwen-Image Lightning LoRA for scenery, Qwen-Image-Edit for characters; group shots text-only with 'different faces' spelled out to avoid face bleed
Wan 2.x I2V (4-step Lightning): 4 min per 7s shot instead of 45-60; gate clips on sharpness vs. the still
InfiniteTalk (4-step) lip-sync, close-ups only, driven by the Hindi track
AI4Bharat Indic Parler-TTS (Mini; Large noisy on ROCm), levelled to -16 LUFS
ACE-Step 1.5 score, ducked 10-12 dB
Real-ESRGAN animevideov3 x2 + Blender depth parallax

Hard-won lessons

Negations in positive prompts get drawn — exclusions go only in the negative prompt
InfiniteTalk animates every visible face; use a calm wide shot plus a separate speaker close-up
480p+upscale distorts eyes; native 720p then ESRGAN x2
Don't crop >1.5x from stills; culture-specific prompts need explicit details
Two ComfyUI servers on one GPU gain only 10-20%; RAM is the real limit

Throughput: 3.5-4 GPU-hours per episode.

Original post →

Fully local open-weight pipeline produces 10-min animated episodes at ~4 GPU-hours each

More from coding & agent