Character Swap LoRA fixes MiniMax H3 video consistency, tests show 88% denoise sweet spot
not_food · reddit · 2026-09-29
Reddit user notfood shares hands-on tests of MiniMax H3 video generation: Ref2VA has known quality issues with mid-distance subjects and colors (acknowledged by the MiniMax team), while the Hybrid model trades video referencing for FL2VA-level quality.
Key takeaways:
- A Character Swap LoRA restores and even improves that lost capability; it also benefits Ref2VA noticeably
- At 88% denoise plus the LoRA, character accuracy is best, though suitability varies per video
- Comparison tests are in the comments; the LoRA and hybrid models are on Hugging Face
More from Multimodal
- NUS rethinks DiT residual connectivity: 1.73x fewer training iterations, 1.39 FID — NationalUniversityofSingapore · 2026-09-29
- Seedance 2.5 makes fake street-footage viral videos: fat cat on scooter prompt shared — LudovicCreator · 2026-09-29
- Open-source ComfyUI enhancer node visibly improves Krea2 turbo 8-step image quality — No-Bad-4838 · 2026-09-29
- ControlNet Depth with end_percent below 0.5 produces surprisingly good results — MaleMaldives · 2026-09-29
- AI-Generated 'Harry Potter and the Order of the GYM' Goes Viral on Reddit — iquizuanswer · 2026-09-29
- Indie dev recreates Manus 2.0-style launch videos with Claude and Seedance — yihui_indie · 2026-09-29