MiniMax H3 swaps characters natively via ref2va — the LoRA only helps with fast motion
arthan1011 · reddit · 2026-09-29
The author notes the community is hyping a new character-swap LoRA, but MiniMax H3's ref2va model already supports character swapping natively — stylized or realistic — without any adapter.
The LoRA still has a use: with hard-to-follow, high-speed reference motion, running it at full strength (1.0) boosts adherence to the ref video. Side-by-side clips with the same prompt are provided.
Related event: MiniMax H3 character swap LoRA tops Hugging Face trending list(4 posts)→
More from Multimodal
- FLUX.2-Klein-Multi-LoRA Space Hits Hugging Face Trending — Hittler-AI · 2026-09-29
- Agentic filmmaking experiment: GPT-6 Astra edits, Grok Imagine renders, Suno v6 scores — TinfoilTricorn · 2026-09-29
- Fizgig 6.6 adds edit-LoRA training for Qwen Image 2.1, learns grading from 40 pairs — shootthesound · 2026-09-29
- Deforum keeps going and going: an endless morphing chair animation — makeitrad1 · 2026-09-29
- GPT Image 2.5 vs open-source Krea 2 Turbo: same-prompt test highlights OpenAI's heavy censorship — Due_Research9042 · 2026-09-29
- Hugging Face release: Yue2-based quiet-storm model for 90s Jazz-Soul vibes — -becausereasons- · 2026-09-29