MiniMax H3 Combined Model Loader
A single ComfyUI loader node for MiniMax H3 FL2VA and REF2VA diffusion models, for combined i2i+ref2i inference using ComfyUi MiniMax H3 Image And Reference To Video node.
MiniMax H3 Combined Model Loader (FL2VA & REF2VA)
A single ComfyUI loader node for MiniMax H3 FL2VA and REF2VA diffusion models useful for purpose of running image to video (i2v with starting and/or ending frame) with references (ref2v).
<img width="1235" height="636" alt="image" src="https://github.com/user-attachments/assets/abae78b3-ccf9-4aee-85c1-45b620074df1" />Modes
- load fl2va+ref2va — default. Builds a combined model on the fly from matching FL2VA and REF2VA checkpoints.
- load fl2va model — loads the selected FL2VA checkpoint using ComfyUI's normal diffusion-model loader path.
- load ref2va model — loads the selected REF2VA checkpoint using ComfyUI's normal diffusion-model loader path.
Combined mode
Combined mode uses the FL2VA checkpoint as the base model and replaces adaln_proj.linear tensors in the selected transformer-block range with the matching tensors from REF2VA.
The default range is 25–49 (inclusive).
The two checkpoints should use matching architectures and quantization formats. For example, use FL2VA and REF2VA versions of the same FP8, INT8 ConvRot, or other matching checkpoint format.
For embedded quantizations such as INT8 ConvRot, leave weight_dtype set to default.
Additionally it features attention override:
- Default (don't override)
- Pytorch (SDPA)
- sageattn_qk_int8_pv_fp16_cuda
- sageattn_qk_int8_pv_fp16_triton
- sageattn_qk_int8_pv_fp8_cuda
- sageattn_qk_int8_pv_fp8_cuda++
- sageattn3
- sageattn3_per_block_mean
(performance note: I got: 300s on SDPA, ~200s on sageattn_qk_int8_pv_fp8_cuda++ which is Sage Attention 2, 185s on sageattn3. If you don't have sageattn3 use SDPA + KJ MiniMax H3 Mem Eff Sage Attention Patch node - 180s. With all tests additionally I also used MiniMax H3 Sparse Attention Turbo Lora, which gives 185s on sageattn3 vs ~200s with ordinary MiniMax H3 Turbo lora)
Installation
Place the folder in:
ComfyUI/custom_nodes/ComfyUI-MiniMax-H3-Combined-Model-Loader/
Then restart ComfyUI.
Node
MiniMax H3 Combined Model Loader
Output: standard ComfyUI MODEL.
Workflow integration
Its purpose is combined I2I+Ref2I workflow.
it might require updating ComfyUI to most recent version (currently v0.33.1)