---
license: apache-2.0
base_model:
- Comfy-Org/MiniMax-H3
---

# MiniMax-H3-Ref-Patch

By extracting 112 keys related to the keywords:  
```
adaln_proj
condition_proj
video_patch_proj
audio_patch_proj
final_layer.audio_out
final_layer.video_out
```
from `ref2va` and `fl2va` weights and storing their differences as patches, we can use ~140MB of data to make fl2va mimic the behavior and output quality of ref2va to some extent.

- You need to use comfyui extension "[ComfyUI-MiniMaxH3_Ref-Patch](https://github.com/lihaoyun6/ComfyUI-MiniMaxH3_Ref-Patch)" to load this patch.
- The goal of this project is not to replicate the behavior of ref2v, but to simulate the effect of **[KJ's LoRA](https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_ref_lora_rank_256_bf16.safetensors)**