README.md
5.2 KB · 110 lines · markdown Raw
1 ---
2 license: other
3 license_name: minimax-h3-community-license-agreement
4 license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
5 base_model: MiniMaxAI/MiniMax-H3
6 base_model_relation: quantized
7 pipeline_tag: image-text-to-video
8 tags:
9 - comfyui
10 - gguf
11 - text-to-video
12 - image-to-video
13 - image-text-to-video
14 - video-to-video
15 - text-to-audio-video
16 - image-to-audio-video
17 - synchronized-audio-video
18 - reference-to-audio-video
19 - multimodal
20 ---
21
22 # MiniMax H3 Pruned GGUF
23
24 <video controls autoplay loop muted style="width: 100%; border-radius: 8px;" src="https://huggingface.co/Abiray/MiniMax-H3-Pruned-GGUF/resolve/main/video/MiniMax_H3_00002-audio.mp4"></video>
25
26 This repository (`Abiray/MiniMax-H3-Pruned-GGUF`) provides pruned and quantized GGUF weights for the **MiniMax H3** omni-modal generative model. MiniMax H3 is designed for unified multimodal context processing, capable of generating synchronized high-definition video and 32 kHz stereo audio from text, image, audio, and video inputs.
27
28 ---
29
30 ## 🌟 Key Highlights
31
32 * **VRAM Efficiency:** Pruned architecture compressed down to **8.9 GB – 21.6 GB**, bringing MiniMax H3 execution to consumer-tier GPUs.
33 * **Synchronized Omni-Modal Output:** Simultaneous generation of video (24 FPS) and native stereo audio (32 kHz).
34 * **Native ComfyUI Support:** Directly compatible with standard `ComfyUI-GGUF` workflows using the native MiniMax backend.
35 * **Dual Pipeline Variants:** Full quant suites for both `FL2VA` (First/Last Frame) and `Ref2VA` (Omni-Reference) modes.
36
37 ---
38
39 ## 📂 Repository Weights & Quantization Breakdown
40
41 > **Note:** For optimal performance, **Q4_K_M** is recommended for 16 GB GPUs, while **Q5_K_M** is recommended for GPUs with 24 GB VRAM.
42
43 ### FL2VA Models (First-and-Last-Frame Mode)
44
45 | Quantization | File Name | Size | Target Hardware / Use Case |
46 | :--- | :--- | :--- | :--- |
47 | **Q3_K_M** | `MiniMax-H3-FL2VA-Pruned-Q3_K_M.gguf` | 8.9 GB | Low VRAM (~12 GB GPUs) |
48 | **Q4_K_M** | `MiniMax-H3-FL2VA-Pruned-Q4_K_M.gguf` | 11.6 GB | **Recommended Balance** (16 GB GPUs) |
49 | **Q4_K_S** | `MiniMax-H3-FL2VA-Pruned-Q4_K_S.gguf` | 11.6 GB | Compact 4-bit K-Quant |
50 | **Q5_K_M** | `MiniMax-H3-FL2VA-Pruned-Q5_K_M.gguf` | 14.1 GB | **High Quality Balance** (24 GB GPUs) |
51 | **Q5_K_S** | `MiniMax-H3-FL2VA-Pruned-Q5_K_S.gguf` | 14.1 GB | Compact 5-bit K-Quant |
52 | **Q6_K** | `MiniMax-H3-FL2VA-Pruned-Q6_K.gguf` | 16.7 GB | Near-Lossless Precision |
53 | **Q8_0** | `MiniMax-H3-FL2VA-Pruned-Q8_0.gguf` | 21.6 GB | Maximum Fidelity / Reference |
54
55 ### Ref2VA Models (Omni-Reference Mode)
56
57 | Quantization | File Name | Size | Target Hardware / Use Case |
58 | :--- | :--- | :--- | :--- |
59 | **Q3_K_M** | `MiniMax-H3-Ref2VA-Pruned-Q3_K_M.gguf` | 8.9 GB | Low VRAM (~12 GB GPUs) |
60 | **Q4_K_M** | `MiniMax-H3-Ref2VA-Pruned-Q4_K_M.gguf` | 11.6 GB | **Recommended Balance** (16 GB GPUs) |
61 | **Q4_K_S** | `MiniMax-H3-Ref2VA-Pruned-Q4_K_S.gguf` | 11.6 GB | Compact 4-bit K-Quant |
62 | **Q5_K_M** | `MiniMax-H3-Ref2VA-Pruned-Q5_K_M.gguf` | 14.1 GB | **High Quality Balance** (24 GB GPUs) |
63 | **Q5_K_S** | `MiniMax-H3-Ref2VA-Pruned-Q5_K_S.gguf` | 14.1 GB | Compact 5-bit K-Quant |
64 | **Q6_K** | `MiniMax-H3-Ref2VA-Pruned-Q6_K.gguf` | 16.7 GB | Near-Lossless Precision |
65 | **Q8_0** | `MiniMax-H3-Ref2VA-Pruned-Q8_0.gguf` | 21.6 GB | Maximum Fidelity / Reference |
66
67 ---
68
69 ## ⚙️ Quickstart & ComfyUI Deployment
70
71 ### Requirements
72 * **ComfyUI:** Version `v0.30.0` or higher is required for native MiniMax-H3 architecture support.
73 * **Extension:** `ComfyUI-GGUF` custom node package installed.
74
75 ### Setup Steps
76 1. Download your desired `.gguf` variant from the table above.
77 2. Place the downloaded `.gguf` file into the `ComfyUI/models/unet/` directory.
78 3. In your ComfyUI workflow, load the model using the **`UnetLoaderGGUF`** node.
79
80 ---
81
82 ## 📋 Model Variants & Input Specifications
83
84 * **H3-Base-FL2VA (First-and-Last-Frame Mode):**
85 * **No image input:** Operates as standard Text-to-Video / Text-to-Audio-Video.
86 * **Single image input:** Generates video conditioned on the initial frame or final frame.
87 * **Dual image input:** Interpolates seamlessly between designated first and last reference frames.
88
89 * **H3-Base-Ref2VA (Omni-Reference Mode):**
90 * **Images:** Up to 9 reference images.
91 * **Videos:** Up to 3 video clips (2–15 seconds per clip; total combined duration $\le$ 15s).
92 * **Audio:** Up to 3 audio clips (must be accompanied by visual input; duration 2–15s).
93 * **Combined Context:** Maximum of 12 multi-modal input assets allowed per prompt.
94
95 ---
96
97 ## 📊 Output Specifications
98
99 * **Duration:** 4 – 15 seconds.
100 * **Aspect Ratios:** Flexible ratio support (16:9, 9:16, 4:3, 3:4, 1:1, 21:9).
101 * **Native Resolution:** 768p base resolution (scalable to 2K via H3-Regenerate-2K).
102 * **Frame Rate:** 24 FPS.
103 * **Audio Track:** 32 kHz stereo audio output.
104 * **Multilingual Dialogue:** Stable support across 11 primary languages (English, Chinese, French, German, Japanese, Korean, Spanish, Russian, Portuguese, Italian, Arabic).
105
106 ---
107
108 ## 📜 Legal & License
109
110 MiniMax H3 is released under the **MiniMax H3 Community License Agreement**. Please refer to the [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) repository and the repository's `LICENSE` file for full terms and commercial usage guidelines.