| --- |
| license: other |
| license_name: minimax-h3-community-license |
| license_link: https://huggingface.co/Comfy-Org/MiniMax-H3 |
| --- |
| |
| # Technical notes |
|
|
| Supporting documentation for |
| [Impact-Flow-v2](https://huggingface.co/yichengup/Impact-Flow-v2). |
|
|
| | | | |
| |---|---| |
| | Base model | MiniMax H3 (`fl2va` family) | |
| | Network | LoRA, rank 16, alpha = rank | |
| | Coverage | all 50 blocks · `attn.qkv_proj`, `attn.out_proj`, `mlp.fc1`, `mlp.fc2` | |
| | Size | 155 MB · 416 tensors | |
| | Training | ai-toolkit 0.13.19 · 1000 steps · 44 epochs | |
|
|
| ## Key naming |
|
|
| Keys use the `diffusion_model.blocks.N.<module>.lora_A / lora_B` convention, so the adapter |
| loads through ComfyUI's standard LoRA path without conversion. |
|
|
| ## Pairing notes |
|
|
| The demo published alongside |
| [minimax_h3-tyc-zs05-ltxqnorm-int8-convrot](https://huggingface.co/yichengup/minimax_h3-tyc-zs05-ltxqnorm-int8-convrot) |
| runs this LoRA together with a 4-step turbo adapter and a sword-combat LoRA. |
|
|
| A fourth adapter in that setup is a block-attenuated derivative of a third-party LoRA, and |
| is **not** redistributed here. For the same look, apply block-gated attenuation to any |
| rank-16 H3 LoRA of your own: keep blocks 35–41 at 0.5 and blocks 42–49 at 0.2. |
|
|
