ComfyUI Node
MiniMax H3 Combined Model Loader
Load MiniMax H3 as FL2VA+REF2VA combined on the fly, or load either FL2VA or REF2VA as a normal standalone diffusion model.
MiniMax H3 Combined Model Loader
- model
◄fl2va_model▾►
◄ref2va_model▾►
◄ref_start_block25►
◄ref_end_block49►
◄weight_dtypedefault►
◄attentiondefault►
◄allow_compilefalse►
◄load_modeload fl2va+ref2va►
Categorymodel/loaders
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| fl2va_model | COMBO | FL2VA checkpoint. Used by 'load fl2va+ref2va' and 'load fl2va model'. | |
| ref2va_model | COMBO | REF2VA checkpoint. Used by 'load fl2va+ref2va' and 'load ref2va model'. | |
| ref_start_block | INT | 250–49 | Combined mode only: first block whose adaln_proj comes from REF2VA (inclusive). |
| ref_end_block | INT | 490–49 | Combined mode only: last block whose adaln_proj comes from REF2VA (inclusive). |
| weight_dtype | COMBO | default | Same weight dtype override as ComfyUI's Load Diffusion Model node. Leave at default for embedded quantizations such as INT8 ConvRot. |
| attention | COMBO | default | Attention backend. 'default' leaves ComfyUI unchanged; 'pytorch' forces the PyTorch/SDPA path. SageAttention modes require their corresponding optional package. SageAttention3 requires supported Blackwell hardware. |
| allow_compile | BOOLEAN | false | Allow torch.compile around SageAttention extension calls. Leave disabled unless your SageAttention setup supports it. |
| load_mode | COMBO | load fl2va+ref2va | Combined mode uses both files and the selected REF2VA block range. Standalone modes ignore the range. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |