ComfyUI AI Dance Video Pipeline Feedback: Identity Drift and Motion Stability
jugernaut126 · reddit · 2026-08-30
The author shared a 35-second vertical dance video prototype generated via ComfyUI using the Minimax H3 model. The video consists of a close-up hook, a profile shot with strobes, and a fast hip-hop section. Key issues identified include face identity drift during head turns (00:07), hand instability under strobe lights (00:12), and a 'floaty' motion feel in the fast section (00:20+). The author seeks technical feedback on identity consistency, temporal stability, and root cause diagnosis within the pipeline (base generation vs. motion control vs. denoising). Technical specs: 768x1376 resolution, 24 FPS, ref2v method, same seed for consistency, with no active motion control or editing applied.
More from Multimodal
- Priming Models: Teacher-Free One-Step Generation Hits FID 1.19 With 12x Less Compute — du_yilun · 2026-10-15
- Same prompt, three takes: Claude vs ChatGPT vs Figma Make UI showdown — Tegadesigns · 2026-10-15
- Designer finds Opus 5.5 picks up nuanced aesthetic taste with startling speed — floguo · 2026-10-15
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15