| --- |
| license: other |
| library_name: diffusers |
| tags: |
| - text-to-image |
| - z-image |
| - diffusers |
| - quantized |
| - int8 |
| - sdnq |
| - safetensors |
| pipeline_tag: text-to-image |
| --- |
| |
| # Tongyi-MAI/Z-Image-Turbo - Quantized (8-bit) |
|
|
| ## Overview |
| This is a **quantized version** of [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo). |
|
|
| All components have been quantized to 8-bit using SDNQ, while preserving the original folder structure for seamless integration. |
|
|
| ## Architecture |
| - **Pipeline**: ZImagePipeline |
| - **Main component**: ZImageTransformer2DModel |
| - **Quantization**: 8-bit |
|
|
| ## Usage |
|
|
| ```python |
| import torch |
| from diffusers import ZImagePipeline, AutoencoderKL, FlowMatchEulerDiscreteScheduler |
| from transformers import Qwen3Model, AutoTokenizer |
| from sdnq import load_sdnq_model |
| |
| model_path = "Tongyi-MAI_Z-Image-Turbo-int8" |
| |
| # Load transformer with SDNQ (quantized to 8-bit) |
| transformer = load_sdnq_model( |
| f"{model_path}/transformer", |
| model_cls=ZImageTransformer2DModel, |
| device="cpu" |
| ) |
| |
| # Load other components from this model (all included!) |
| vae = AutoencoderKL.from_pretrained(f"{model_path}/vae", torch_dtype=torch.float16) |
| text_encoder = Qwen3Model.from_pretrained(f"{model_path}/text_encoder", torch_dtype=torch.float16) |
| tokenizer = AutoTokenizer.from_pretrained(f"{model_path}/tokenizer") |
| scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(f"{model_path}/scheduler") |
| |
| # Construct pipeline |
| pipe = ZImagePipeline( |
| transformer=transformer, |
| vae=vae, |
| text_encoder=text_encoder, |
| tokenizer=tokenizer, |
| scheduler=scheduler, |
| ) |
| |
| pipe.to("cuda") |
| |
| # Generate an image |
| image = pipe( |
| prompt="A serene mountain landscape at sunrise", |
| num_inference_steps=20, |
| ).images[0] |
| image.save("output.png") |
| ``` |
|
|
| ## Components |
|
|
| - ✅ **transformer** (ZImageTransformer2DModel) - Quantized to 8-bit |
| - ✅ **vae** (AutoencoderKL) - Quantized to 8-bit |
|
|
| **Note**: Some components are included unquantized due to SDNQ library limitations: |
| - 📦 **text_encoder** - Included unquantized (SDNQ bug workaround) |
| |
| |
| ## Quantization Details |
| - **Original model**: [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) |
| - **Quantization**: 8-bit |
| - **Quantizer**: SDNQ |
| - **Date**: 2026-01-18 13:41:50 |
| |
| ## Size Reduction |
| - Original: ~30GB (estimated) |
| - Quantized: See individual component sizes |
| |
| ## Notes |
| - This is a complete drop-in replacement - all components included |
| - SDNQ quantization provides excellent quality at reduced size |
| - Requires `sdnq` library to be installed: `pip install sdnq` |
| - Quality loss is minimal with 8-bit quantization |
| - Some components may be included unquantized due to library limitations |
| |
| --- |
| Quantized with [BugQuant](https://github.com/yourusername/BugQuant) |
| |
