Anime-to-live-action with just a video reference + LoRA: MiniMax H3 test beats expectations
aiyakisoba · reddit · 2026-10-13
- A Reddit user tested anime-to-realism conversion in MiniMax H3: no keyframes or stills, just feeding the original anime video as a video reference, applying a community LoRA (ref2vAnime2Realism on Hugging Face), and prompting for live-action style.
- Output quality was better than expected, though not perfect; without a character reference sheet, cross-scene character consistency isn't guaranteed and depends on how the model interprets the character per scene.
- Next step: adding a character reference sheet to improve consistency. Author is open to questions.
More from Multimodal
- Runway's best month yet: one General World Model backbone powers four product lines — tlakomy · 2026-10-13
- One Prompt Gets Claude Opus 5.5 to Generate a Full Beat-Synced K-pop Music Video — minchoi · 2026-10-13
- Open-source ff-tracking shines for kinetic typography lyric videos, creator says — PurzBeats · 2026-10-13
- Surreal AI-generated Monica Bellucci fashion images with giant hands go viral — OVolosin82152 · 2026-10-13
- Sarvam AI launches Bulbul v4 Flash, a faster TTS for Indian languages — itsOmSarraf_ · 2026-10-13
- Scene Studio: one prompt turns your scene into 100 animated art styles — mhmazur · 2026-10-13