RefMod for MiniMax H3 tested: 22 images to a reusable character, faster than LoRA
Sad_Coach_1433 · reddit · 2026-09-07
A first hands-on test of RefMod, a new character-consistency feature for MiniMax's H3 video model: pack 22 reference images of a character into a .safetensors file, then select it from a list in a custom node to keep the same character across videos — no need to load images each time, and faster than LoRA training. An installation and usage guide is available on Hugging Face. The main drawback: it only preserves appearance, not the character's voice, so you still need a separate audio reference for speech.
More from Multimodal
- One person made a 94-minute AI sci-fi film adapting Liu Cixin's 'Mountain' for ~$28k RMB — xiaosun86 · 2026-09-07
- FastH3 + USDU doubles video upscale speed on RTX 3090: 48m vs 96m, full workflow released — inazma44 · 2026-09-07
- 'Mirror That Shows You 10 Years Older' AI Video Wows Reddit — Weak-Abalone-9647 · 2026-09-07
- Astra turns photogrammetry scan of old home into editable 3D scene, hinting inverse graphics has crossed the chasm — bilawalsidhu · 2026-09-07
- Trellis2/pixel3d ComfyUI template workflow generates detailed 3D models in minutes — Disastrous-Agency675 · 2026-09-07
- "It's the artist that doesn't need the studios": Creator demos Minimax H3 workflow in ComfyUI — jnack · 2026-09-07