RefMod Without LoRA Training: Bundling a Character, a Fight Move, a Voice and a Set Into Loadable Files
The usual way to make a model hold a specific character or perform a specific fight move is to train a LoRA. RefMod does something different: it stores the reference as a loadable modification file that conditions the mo
The usual way to make a model hold a specific character or perform a specific fight move is to train a LoRA. RefMod does something different: it stores the reference as a loadable modification file that conditions the model at inference time.
The practical difference is turnaround. A LoRA takes a training run; a RefMod takes the source material and the extraction pass.
What goes into a RefMod
Six workflows in this episode, in four categories:
| # | Workflow | Reference type |
|---|---|---|
| 01 | Character RefMod | Single image or folder batch |
| 02 | Action RefMod | Single video or folder batch |
| 03 | Audio RefMod | Single audio clip |
| 04 | Martial-arts-hall scene RefMod | Two camera angles, batch |
| 05 | Multi-reference gen A — Apply | Applies a stack of RefMods |
| 06 | Multi-reference gen B — TextEncode | Encodes the textual half of the conditioning |
The A/B split in 05 and 06 matters: applying a reference and describing it are two separate stages, and you can change one without rebuilding the other.
Where it is actually strong
- Character: holds identity across shots. Face structure, hair, and the read of the outfit survive camera changes.
- Action: a fight move transfers as motion, not as a still pose. That is the hard part and the reason this is interesting.
- Scene: a fixed set with two camera angles means you can cut between coverage without the room changing.
Where it is weaker
Audio RefMod is the least reliable of the four. Timbre transfers, but the cadence of the source clip leaks into the output — if the reference was recorded quickly, the result sounds rushed regardless of what the text says.
The realistic workflow
Use RefMod to get 90% there, then fix the remaining 10% with a short LoRA only if you need it. That order saves the training run in most cases.
Files
- HF: https://huggingface.co/AI-KSK/minimax-h3-refmod-workflows
- ModelScope: https://modelscope.cn/models/Sunjian520/minimax-h3-refmod-workflows
Video walkthrough (Chinese): https://youtu.be/H5L5sO7dRu8
Bilibili: https://www.bilibili.com/video/BV1F2nZ6SEA7/
Notes
- ComfyUI exports; the H3 model weights must exist locally.
- Tutorial material, not a product.
- Independent AI-KSK release; not affiliated with the model authors or RunningHub.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.