Asirus/TaoMate_H3_3_Step_LoRA

TaoMate-H3 3-Step LoRA (ComfyUI)

Official TaoMate-H3 3-step LoRA from Alibaba TaoLive AIGC Team, converted for ComfyUI.

Allows generating video + audio with MiniMax H3 in just 3 steps.

---

Quality Comparison

Turbo 4-step v1.2	TaoMate MAXQUALITY (BF16)	TaoMate LIGHT (FP16)
▶️ Watch Video	▶️ Watch Video	▶️ Watch Video

---

Files

File	Rank	Precision	Size	Recommendation
taomate_h3_3step_MAXQUALITY_bf16.safetensors	128	BF16	1.2 GB	Best Quality
taomate_h3_3step_LIGHT_fp16.safetensors	128	FP16	1.2 GB	Good Alternative

---

Recommended Settings

Parameter	Value
Steps	3
CFG	1.0
Sampler	euler or er_sde
Scheduler	simple
Strength	0.9 – 1.0
Video Shift	8–12
Audio Shift	3

Works best with the FL2VA version of MiniMax H3.

---

Benchmarks

All tests were performed with the same prompt, 5 seconds duration, and 1 MP resolution.

6 GB VRAM

Model + LoRA	Steps	Generation Time
FL2VA pruned int8 + Turbo 4-step v1.2	4	474.30 sec
FL2VA pruned int8 + TaoMate MAXQUALITY	3	373.22 sec
FL2VA pruned int8 + TaoMate LIGHT	3	380.59 sec

16 GB VRAM

Model + LoRA	Steps	Generation Time
FL2VA pruned int8 + Turbo 4-step v1.2	4	130.51 sec
FL2VA pruned int8 + TaoMate MAXQUALITY	3	102.81 sec
FL2VA pruned int8 + TaoMate LIGHT	3	104.84 sec

Conclusion: TaoMate 3-step is approximately 21–27% faster than Turbo 4-step while delivering similar or better quality.

---

Conversion Details

Original model: TaoLiveAIGC/TaoMate-H3

What was done:

• Keys renamed from blocks.*.lora_a / lora_b → diffusion_model.blocks.*.lora_down.weight / lora_up.weight
• Added diffusion_model. prefix
• Precision converted to BF16 and FP16
• No fine-tuning and no merging with the base model

---

Hardware Recommendations

• 6–8 GB → Both versions work (prefer MAXQUALITY)
• 10 GB and above → Use MAXQUALITY_bf16

---

Original Model & License

• Original LoRA: TaoLiveAIGC/TaoMate-H3
• Base model: MiniMaxAI/MiniMax-H3
• License: MiniMax H3 Community License

Converted for the ComfyUI community.