slmonker/TaoMate-HD-preview-comfy

TaoMate-HD Preview for ComfyUI

Converted ComfyUI-compatible LoRA weights for the TaoMate-HD three-stage progressive sampler with MiniMax H3 audio-video generation.

This repository contains the converted adapter used by:

• GitHub: https://github.com/slmonker/ComfyUI-TaoMateHD
• Upstream TaoMate-HD: https://github.com/TaoLiveAIGC/TaoMate-HD
• Base model: https://huggingface.co/MiniMaxAI/MiniMax-H3

File

TaoMate-HD-preview-comfy.safetensors

• Format: ComfyUI-compatible MiniMax H3 LoRA
• Precision: FP32
• Size: 2,480,994,680 bytes (approximately 2.48 GB)
• Modules: 208
• Tensors: 624 (lora_A, lora_B, and per-module alpha)
• SHA-256: 352a9ace5c5f4e99d52536267259a000eb514cbb62b745942bb3768cc0df6ec4

The conversion renames the upstream lowercase lora_a / lora_b keys to ComfyUI's lora_A / lora_B format and writes TaoMate's per-module alpha value (128) so the intended LoRA scaling is preserved.

Do not load the original upstream adapter_model.safetensors directly in ComfyUI.

Download

hf download slmonker/TaoMate-HD-preview-comfy \
  TaoMate-HD-preview-comfy.safetensors \
  --local-dir ComfyUI/models/loras

Use in ComfyUI

LoraLoaderModelOnly
  lora_name = TaoMate-HD-preview-comfy.safetensors
  strength_model = 1.0

Recommended sampling pipeline:

MiniMax H3 model
  -> LoraLoaderModelOnly (strength_model=1.0)
  -> MiniMaxH3SigmaShift (shift_video=12, shift_audio=3)
  -> TaoMate HD Progressive Sampler
  -> Separate AV Latent
       -> VAEDecode
       -> VAEDecodeAudio

The ComfyUI adapter supports:

• T2VA
• Ref2VA through native MiniMax H3 minimax_refs conditioning, experimental
• Three-stage TaoMate sampling with steps=3

The adapter does not currently implement FL2VA first/last-frame keyframe conditioning or denoise masks.

Compatibility

Use a MiniMax H3 checkpoint with the required task capability:

• T2VA uses the native MiniMax H3 image-to-video node without first/last frames.
• Ref2VA requires a Ref2VA-capable or hybrid H3 checkpoint and native MiniMax H3 Reference to Video conditioning.
• The LoRA should be applied at strength 1.0.
• An int8/fp8 H3 checkpoint is recommended for 32 GB GPUs.

Attribution

This is an unofficial community conversion. It is not an official release from Alibaba TaoLive AIGC or MiniMax.

The adapter and base-model use remain subject to the MiniMax H3 Community License Agreement. See LICENSE and NOTICE.