CZMartin22/TaoMate-H3-3step-ComfyUI

TaoMate-H3 3-Step LoRA (ComfyUI Format, BF16)

This repository provides a ComfyUI-compatible conversion of the official TaoLiveAIGC/TaoMate-H3 step-3000 EMA LoRA adapter for MiniMax H3 (FL2VA).

The original weights were released in PEFT / Diffusers format with non-standard key naming. This converted version maps all weights directly to ComfyUI's DiT structure, embeds the required alpha scalar tensors, and casts the weights to bfloat16 to save VRAM and disk space (~1.24 GB vs original ~2.48 GB).

When loaded in ComfyUI, all 208 patches attach cleanly without missing key warnings.

---

Model Details

• Base Model: MiniMax H3 (FL2VA)
• Original Adapter: TaoLiveAIGC/TaoMate-H3 by Alibaba TaoLive AIGC Team
• Rank / Alpha: r = 128, alpha = 128
• Format: Safetensors (bfloat16)
• Compatibility: ComfyUI (standard LoraLoader / LoraLoaderModelOnly, Pixaroma H3 nodes, ComfyUI-H3)

---

Recommended ComfyUI Settings

Parameter	Recommended Value	Notes
Steps	3	Model is distilled specifically for 3-step generation
CFG	1.0	CRITICAL: Do NOT use CFG > 1.0 (destilled ODE burns with guidance)
LoRA Strength	1.0
Sampler	Euler
Scheduler	simple / linear FlowMatch
Base Model	MiniMax H3 FL2VA	Supports Text-to-Video & Image-to-Video with synchronized Audio

Expected Console Output

[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 208 patches attached.

---

Quick Download via CLI

Place TaoMate-H3-3step-ComfyUI.safetensors into your ComfyUI/models/loras/ directory:

# Using huggingface-cli
huggingface-cli download CZMartin22/TaoMate-H3-3step-ComfyUI TaoMate-H3-3step-ComfyUI.safetensors --local-dir ComfyUI/models/loras/

---

Conversion Details

The conversion script performed the following transformations:

1. Stripped PEFT base_model.model. namespaces.
2. Formatted down/up weights from .lora_a / .lora_b to standard ComfyUI .lora_A.weight and .lora_B.weight.
3. Prepended diffusion_model. to transformer blocks and token refiners.
4. Embedded individual alpha scalar tensors (alpha = 128.0) per module to ensure correct automatic scaling in ComfyUI.
5. Cast FP32 weights to BF16.

---

Credits & License

• Original TaoMate-H3 developed by Alibaba TaoLive AIGC Team.
• Base architecture and weights by MiniMax AI.
• Released under the terms of the MiniMax H3 Community License Agreement.