Fix face loss in MMH3 long shots with prompt timing
MarekNowakowski · reddit · 2026-08-20
To fix face and scene loss in MMH3 videos longer than 10 seconds, the author found that inserting the reference image tag in the middle of the prompt refreshes the model's memory. For example, using "woman from <picture 1> looks left" at second 7 helped retain the face perfectly at second 14 with 0.9 resolution.
More from Multimodal
- FLUX 3 Upscaler hits Higgsfield: 50% faster and cheaper, native 2K/4K video — tisch_eins · 2026-10-16
- AI-generated "RiRi" images stun with photorealistic diamond-gown Rihanna looks — egeberkina · 2026-10-16
- SeedAudio 1.5 hits CapCut PC: one prompt generates dialogue, SFX and music — LearnWithBishal · 2026-10-16
- DLSS 5 Assetto Corsa demo stuns sim racers: 'better than reality' — CSProfKGD · 2026-10-16
- Video recast with fal agent: swap a character via one character sheet — adamho · 2026-10-16
- Dev open-sources Krea 2 pixel-art LoRA for 16x16 game icons with transparent backgrounds — Angaisb_ · 2026-10-16