---

language:

• en license: other license_name: minimax-license license_link: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/LICENSE tags:
• text-to-video
• image-to-video
• video-generation
• minimax-h3
• comfyui
• quantized
• int8
• convrot
• dmad
• pdmd
• distilled
• accelerated base_model:
• Comfy-Org/MiniMax-H3
• pdmd2026/pdmd_4NFE_lora
• ZhengmingYu/DMAD pipeline_tag: text-to-video

---

Minimax-H3-Hybrid-DMAD-PDMD (Pruned INT8 ConvRot)

Minimax-H3-Hybrid-DMAD-PDMD is an accelerated, low-VRAM, reference-capable 33B video-generation foundation model built for ComfyUI.

It resolves the long-standing community trade-off between MiniMax-H3's FL2VA (clean dynamic range, natural motion, but missing multi-reference conditioning) and Ref2VA (reference following, but prone to crushed shadows and dark banding) by implementing a Smooth S-Curve Hybrid Merge. It then permanently folds two state-of-the-art distillation LoRAs—DMAD 8-step and PDMD 4-step—into the transformer weights and compiles the entire model into the Pruned INT8 ConvRot format (~18.89 GB, down from >61 GB BF16).

---

🌟 Highlights & Technical Innovations

1. Dual-Distillation Superposition (4–8 Steps):
   • Fuses DMAD (Distribution Matching Adversarial Distillation) at 0.85 strength to enforce high-frequency critic texture retention.
   • Fuses PDMD (Projected Distribution Matching Distillation) at 0.333 strength to project out student–critic endpoint residuals, stabilizing fast trajectory convergence down to 4 NFE.
2. Smooth S-Curve Hybrid AdaLN Injection:
   • Blocks 0–19 (100% FL2VA): Guarantees pristine contrast, realistic skin tones, dynamic lighting, and unclipped black levels.
   • Blocks 20–34 (Cubic Smoothstep Transition): Progressively introduces Ref2VA modulation via $S(t) = t^2(3 - 2t)$ to eliminate inter-block latent shock.
   • Blocks 35–49 (100% Ref2VA): Delivers deep multimodal reference conditioning (subject image, reference video, audio guidance).
   • Final Layer: 50/50 balance protecting output color space while locking subject identity.
3. Pruned INT8 ConvRot Architecture:
   • Rank-8 Time Modulation Pruning: The $2688$-dimensional dense timestep projections are compressed into an $8$-dimensional subspace via canonical SVD affine folding ($W \cdot \text{basis}^T \in \mathbb{R}^{96768 \times 8}$, $b + W \cdot \text{mean} \in \mathbb{R}^{96768}$), accompanied by the lookup grid adaln_t_table [1025, 8].
   • Orthogonal Hadamard Rotation ($H_{256}$): Pre-multiplies weight blocks by normalized $256 \times 256$ Sylvester matrices, suppressing activation outliers and enabling symmetric INT8 tensor-core execution via comfy-kitchen.

---

🔬 Mathematical Formulations

1. Adaptive S-Curve Block Merging

For each transformer block $i \in {0, \dots, 49}$ and its AdaLN modulation parameter $W^{(i)}$:

$$\alpha(i) = \begin{cases} 0.0 & \text{if } i < 20 \ t^2(3 - 2t) \quad \text{where } t = \frac{i - 20}{35 - 20} & \text{if } 20 \le i < 35 \ 1.0 & \text{if } i \ge 35 \end{cases}$$

$$W^{(i)}{\text{hybrid}} = (1 - \alpha(i)) \cdot W^{(i)}{\text{FL2VA}} + \alpha(i) \cdot W^{(i)}_{\text{Ref2VA}}$$

2. Dual-Distillation Weight Superposition

For every linear projection in the 50 DiT blocks and 2 Token Refiner blocks (208 total modules):

$$W_{\text{fused}} = W_{\text{hybrid}} + 0.85 \cdot \left( B_{\text{DMAD}} A_{\text{DMAD}} \cdot \frac{\alpha_1}{r_1} \right) + 0.333 \cdot \left( B_{\text{PDMD}} A_{\text{PDMD}} \cdot \frac{\alpha_2}{r_2} \right)$$

Note on PDMD SwiGLU Alignment: The PDMD LoRA was integrated using the native ComfyUI layout conversion, resolving Diffusers' $[value; gate]$ row order to match ComfyUI's $[gate; up]$ SwiGLU convention.

3. Hadamard ConvRot INT8 Quantization

Linear weights $W \in \mathbb{R}^{N \times K}$ (where $K \pmod{256} = 0$) are rotated prior to row-wise quantization using the Kronecker power of the basic Hadamard matrix $H_4$:

$$H_4 = \frac{1}{2} \begin{pmatrix} 1 & 1 & 1 & -1 \ 1 & 1 & -1 & 1 \ 1 & -1 & 1 & 1 \ -1 & 1 & 1 & 1 \end{pmatrix}, \quad H_{256} = H_4^{\otimes 4}$$

$$W_{\text{rot}} = \left( W.\text{view}\left(N, \frac{K}{256}, 256\right) \cdot H_{256}^T \right).\text{reshape}(N, K)$$

$$s_i = \max\left( \frac{\max_j |W_{\text{rot}, ij}|}{127.0}, , 10^{-30} \right), \quad Q_{ij} = \text{clamp}\left( \left\lfloor \frac{W_{\text{rot}, ij}}{s_i} \right\rceil, -128, 127 \right) \in \text{int8}$$

---

📊 Lineage & Attribution

This model is a synthesis of foundational research and open weights released by the following projects:

Component	Source Model / Artifact	Original Authors / Citations	Role in Hybrid
Base Model 1	minimax_h3_fl2va_bf16	MiniMax AI & Comfy-Org	Primary spatial/motion backbone (Blocks 0–34).
Base Model 2	minimax_h3_ref2va_bf16	MiniMax AI & Comfy-Org	Deep reference-guidance AdaLN pathways (Blocks 20–49).
LoRA 1	dmad_minimax_h3_8step_full_critic_comfyui	Zhengming Yu et al.(DMAD for Fast Video Generation)	High-frequency detail preservation @ 0.85 strength.
LoRA 2 (Original)	pdmd2026/pdmd_4NFE_lora	Zimo Wang, Junkun Yuan, Angtian Wang et al.(arXiv:2609.35768)	Projected Distribution Matching Distillation algorithm.
LoRA 2 (Adapter)	minimax_h3_pdmd_4nfe_comfyui_v6	Iwannapose	ComfyUI SwiGLU-remapped adapter @ 0.333 strength.
Pruning Geometry	adaln_affine.safetensors	Pedro Cuenca (multimodalart)	Canonical SVD basis/mean for low-rank AdaLN folding.

---

⚙️ Recommended ComfyUI Inference Parameters

Because this model contains fused distillation vectors, do not use standard 30–50 step settings or high CFG values, which will cause severe over-saturation.

Steps: 6 to 8 (Optimal) | 4 steps (Fast preview) | 10 steps (Maximum texture fidelity)
CFG (Guidance Scale): 1.0 to 1.5 (Strict: Do NOT exceed 2.0)
Sampler: euler or euler_ancestral (Recommended) | uni_pc / res_multistep (Fast)
Scheduler: simple or sgm_uniform | beta
Denoise: 1.0 (Text-to-Video / First-Frame generation)
Shift: 12.0 / 3.0 (Base MiniMax scheduler default)

---

🛠️ Step-by-Step ComfyUI Setup

1. Directory Structure

Download Minimax-H3-Hybrid-DMAD-PDMD_int8.safetensors and place the required auxiliary models into your ComfyUI root:

ComfyUI/
├── models/
│   ├── diffusion_models/
│   │   └── Minimax-H3-Hybrid-DMAD-PDMD_int8.safetensors  # <-- THIS MODEL (18.89 GB)
│   ├── vae/
│   │   └── minimax_h3_video_vae_fp16.safetensors
│   └── text_encoders/
│       └── qwen2.5_vl_7b_clip_fp16.safetensors          # Or Qwen3-VL text encoder

2. Node Configuration

1. Load Diffusion Model: Select Minimax-H3-Hybrid-DMAD-PDMD_int8.safetensors.
   • ComfyUI automatically detects the internal metadata marker: {"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256} and loads hardware-accelerated kernels via comfy-kitchen.
2. Dual VAE Loader: Load MiniMax Video VAE and Audio VAE.
3. Clip/Text Loader: Connect your text encoder (e.g. Qwen2.5-VL / Qwen3-VL).
4. KSampler: Set steps: 8, cfg: 1.0, sampler_name: euler, scheduler: simple.

---

📈 VRAM Footprint & Speed Benchmarks

Hardware Platform	Precision	Weight Footprint	Generation Steps	Average Latency (720p, 5s)
NVIDIA RTX 4090 (24GB)	BF16 (Dense)	~61.7 GB (OOM - requires CPU swap)	30 steps	~240s
NVIDIA RTX 4090 (24GB)	INT8 ConvRot (This Model)	18.89 GB (Fits entirely in VRAM)	8 steps	~24s (10x faster)
NVIDIA RTX 3060 (12GB)	INT8 ConvRot (This Model)	Low-VRAM Block-streaming	6 steps	~68s
Apple M2/M3 Max (64GB)	INT8 ConvRot (This Model)	18.89 GB Unified Memory	8 steps	~48s

---

🔍 Troubleshooting & Artifact Avoidance

• Plastic skin or harsh outlines? Your CFG is set too high. Reduce CFG to 1.0 or 1.2. Distillation student critics function optimally under minimal classifier-free guidance.
• Jerky motion or camera warping? Increase steps from 4 to 8. While PDMD supports 4 NFE, DMAD's critic benefits significantly from 6–8 integration steps.
• Missing Audio or Video out? Ensure both the video patch projection and audio patch projection are connected in your ComfyUI workflow; this model retains joint video-audio synthesis pathways.

---

📜 Academic Citations

If you use this model or its merge methodology in your research or applications, please cite the foundational works:

@misc{wang2026pdmdprojecteddistributionmatching,
  title={PDMD: Projected Distribution Matching Distillation for Video Diffusion Models}, 
  author={Zimo Wang and Junkun Yuan and Angtian Wang and Haotian Yang and Canyu Zhang and Siyuan Yuan and Xingchang Huang and Bo Liu and Yizhi Wang and Yiding Yang and Chongyang Ma and Gordon Guocheng Qian},
  year={2026},
  eprint={2609.35768},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.35768}
}

@article{yu2026dmad,
  title={Distribution Matching Adversarial Distillation for Fast Video Generation},
  author={Zhengming Yu and others},
  journal={arXiv preprint},
  year={2026}
}

@misc{minimax2025h3,
  title={MiniMax-H3 (Hailuo) Joint Audio-Video Diffusion Model},
  author={MiniMax AI Team},
  year={2025},
  publisher={Comfy-Org},
  howpublished={\url{https://huggingface.co/Comfy-Org/MiniMax-H3}}
}