File size: 4,268 Bytes
4fdac1c ef3f1f8 4fdac1c ef3f1f8 4fdac1c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 |
---
license: other
license_name: tencent-hunyuan-community
license_link: https://huggingface.co/tencent/HunyuanImage-3.0/blob/main/LICENSE.txt
tags:
- text-to-image
- hunyuan
- quantization
- int8
- comfyui
- custom nodes
- autoregressive
- Dit
- Hunyuan-Image-3
pipeline_tag: text-to-image
---
# Hunyuan Image 3.0 - INT8 Quantized
This is an **INT8 quantized version** of Tencent's [HunyuanImage-3.0](https://huggingface.co/tencent/HunyuanImage-3.0) model, optimized for high-end GPU workflows without CPU offloading.
## Model Description
INT8 quantization of the Hunyuan Image 3.0 text-to-image diffusion transformer, providing a balance between the full BF16 precision and more aggressive NF4 quantization. This version maintains excellent image quality while reducing memory requirements.
**Key Features:**
- 🎯 High quality output comparable to BF16
- 💾 ~80GB VRAM required (fits RTX 6000 Ada/Blackwell)
- ⚡ ~3.5 minutes generation time at base resolution
- 🔧 Designed for ComfyUI workflows
## VRAM Requirements
| Phase | VRAM Usage |
|-------|------------|
| Weight Loading | ~80 GB |
| Inference (additional) | ~12-20 GB |
| **Total** | **~92-100 GB** |
**Recommended Hardware:**
- NVIDIA RTX 6000 Ada (48GB) - requires model split/offload
- NVIDIA RTX 6000 Blackwell (96GB) - fits entirely in VRAM ✅ Workflows on the github page
- Multi-GPU setups with 80GB+ combined VRAM
## Usage
### ComfyUI (Recommended)
This model is designed to work with the [Comfy_HunyuanImage3](https://github.com/EricRollei/Comfy_HunyuanImage3) custom nodes:
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
```
Install the nodes and download this model to your ComfyUI models directory. The nodes handle INT8 loading automatically.
### Direct Usage
```python
# INT8 weights can be loaded with standard torch quantization
# See the ComfyUI nodes for reference implementation
```
## Performance
- **Generation Time**: ~3.5 minutes for base resolution (1024x1024)
- **Weight Loading**: ~60 seconds (one-time per session)
- **Quality**: Excellent - minimal degradation from BF16
- **Speed**: Faster inference than BF16 due to reduced memory bandwidth
## Quantization Details
- **Method**: INT8 per-channel quantization
- **Target**: Hunyuan Image 3.0 transformer backbone
- **Precision Loss**: Minimal - image quality remains high
- **Trade-off**: Middle ground between NF4 (lower quality) and BF16 (highest VRAM)
## Original Model
This is a quantized derivative of [Tencent's HunyuanImage-3.0](https://huggingface.co/tencent/HunyuanImage-3.0).
**Original Model Details:**
- Architecture: Diffusion Transformer
- Resolution: Up to 2048x2048
- Language Support: English and Chinese prompts
- License: [Tencent Hunyuan Community License](https://huggingface.co/tencent/HunyuanImage-3.0/blob/main/LICENSE.txt)
Please review the original model card and license for full details on capabilities and restrictions.
## Limitations
- Requires high-end professional GPU (80GB+ VRAM)
- Not suitable for consumer GPUs (4090, 5090) without further optimization
- INT8 quantization may introduce minor quality differences in edge cases
- Loading time adds ~1 minute overhead to first generation
## Credits
**Original Model**: [Tencent Hunyuan Team](https://huggingface.co/tencent)
**Quantization**: Eric Rollei
**ComfyUI Integration**: [Comfy_HunyuanImage3](https://github.com/EricRollei/Comfy_HunyuanImage3)
## License
This model inherits the license from the original Hunyuan Image 3.0 model:
- **License**: [Tencent Hunyuan Community License](https://huggingface.co/tencent/HunyuanImage-3.0/blob/main/LICENSE.txt)
- Please review the original license for commercial use restrictions and requirements
## Citation
```bibtex
@misc{hunyuan-image-3-int8,
author = {Rollei, Eric},
title = {Hunyuan Image 3.0 INT8 Quantized},
year = {2024},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/[YOUR_USERNAME]/[MODEL_NAME]}}
}
```
Original model citation:
```bibtex
@misc{tencent2024hunyuan,
title={Hunyuan Image 3.0},
author={Tencent Hunyuan Team},
year={2024},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/tencent/HunyuanImage-3.0}}
}
``` |