# Z_image_Shaker

![image](https://cdn-uploads.huggingface.co/production/uploads/67f4c86b6cac98d20c31720c/QStVgISZMh2mvRG-_Sgfx.png)

A high-quality NSFW image generation model based on ZImageTurbo architecture.

## Model Files

| File | Size | Precision | Recommended |
|------|------|-----------|-------------|
| `z_image_H_shaker.safetensors` | 22.24 GB | FP32 (Full) | Yes | merge model full meta
| `z_image_E_shaker.safetensors` | 22.24 GB | FP32 (Full) | Yes |
| `z_image_E_shaker_fp8.safetensors` | 5.77 GB | FP8 | No |

## Important Notice

**It is NOT recommended to use the FP8 quantized version.**

The FP8 version (`z_image_E_shaker_fp8.safetensors`) has significant precision loss in the head layers, which results in reduced high-detail performance. The quantization affects the model's ability to generate fine details accurately.

**Only use the FP8 version if:**
- You have limited VRAM (less than 24GB)
- You prioritize speed over quality
- You want a quick preview before downloading the full model

**For best results, always use the full FP32 version** (`z_image_E_shaker.safetensors`).

## Usage

### Recommended Settings
- **Sampler:** sa_solver_beta
- **Scheduler:** sa_solver_beta
- **Keep prompts minimal** - too many prompts will pollute the image quality

### Tips
- Less is more - simple prompts work best
- This model performs excellently on GPUs with 32GB+ VRAM (e.g., RTX 5090)
- The ZImageTurbo architecture provides efficient inference performance

## License

Apache 2.0

## Author

eddy (spaceai2077)