Fixing AI video face melt: upload raw audio and quote dialogue in prompt
Fragrant-Cheek-4273 · reddit · 2026-08-31
The author encountered face melting in AI videos using Minimax H3, attributing it to independent VAEs for audio and visual data. Audio data is less dense, causing the audio encoder to overfit and pull spatial anchors out of alignment.
Fix: upload the raw audio reference file and type the exact dialogue in double quotation marks inside the positive text prompt. This forces the model to map lip movements to the literal text string while referencing the audio VAE, stabilizing facial expressions. Not perfect, but passable.
More from Multimodal
- Orchestrating a 100-image portrait of humanity for aliens with ChatGPT, Argon and p5.brush — dx8xb · 2026-10-16
- Indie Creator Builds an AI-Native Animated Show Solo — ComprehensiveLime654 · 2026-10-16
- Ideogram teases 4.5 arrival and open weights: 'Not yet, but soon' — East_Shoe8814 · 2026-10-16
- Opus 5.5 turns heads with surprisingly polished vertical video demo — FinanceYF5 · 2026-10-16
- 'Celestial Ocean Reverie' prompt template blends starry skies with dreamlike seascapes — LudovicCreator · 2026-10-16
- 'Thicc Lil' Red': an AI-made musical meme reimagining Little Red Riding Hood — bennash · 2026-10-16