| --- |
| license: apache-2.0 |
| base_model: |
| - Tongyi-MAI/Z-Image-Turbo |
| pipeline_tag: text-to-image |
| library_name: diffusers |
| tags: |
| - comfyui |
| - w4a8 |
| --- |
| |
| # Z-Image-Turbo W4A8 |
|
|
| W4A8 (4-bit weight, 8-bit activation) quantized weights for |
| [Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo), made for |
| [ComfyUI](https://github.com/comfyanonymous/ComfyUI) using ComfyUI's native |
| `asym_w4a8_int8` quantized-diffusion format (Comfy Kitchen). |
|
|
| Both the diffusion model and the Qwen3-4B text encoder are quantized, so the |
| whole pipeline fits in low VRAM. |
|
|
| ## Files |
|
|
| | File | Size | Notes | |
| | --- | --- | --- | |
| | `z_image_turbo_w4a8.safetensors` | 3.5 GB | Diffusion model, `asym_w4a8_int8`, group_size 16 + ConvRot | |
| | `qwen_3_4b_w4a8.safetensors` | 2.8 GB | Qwen3-4B text encoder, `asym_w4a8_int8`, group_size 16 + ConvRot | |
| |
| Original BF16 sizes: diffusion 12.3 GB, text encoder 8.0 GB. |
| |
| ## Usage (ComfyUI) |
| |
| Place the files in your ComfyUI `models` directory: |
| |
| ``` |
| ComfyUI/ |
| ├── models/ |
| │ ├── diffusion_models/ |
| │ │ └── z_image_turbo_w4a8.safetensors |
| │ ├── text_encoders/ |
| │ │ └── qwen_3_4b_w4a8.safetensors |
| │ └── vae/ |
| │ └── flux1-vae.safetensors |
| ``` |
| |
| Then use the standard Z-Image-Turbo text-to-image workflow with a `Load Diffusion |
| Model` node pointed at `z_image_turbo_w4a8.safetensors` and a `Load CLIP` node |
| pointed at `qwen_3_4b_w4a8.safetensors`. |
|
|
| Both files are detected automatically by ComfyUI (`.comfy_quant` metadata keys); |
| no custom nodes are required. The text encoder must be loaded through the |
| Qwen3-4B / Z-Image CLIP path (it does not need the pooled output). |
|
|
| ## Quality & Speed |
|
|
| Verified on an 8 GB VRAM GPU (RTX 4060 Laptop) at 1024x1024, 8 sampling steps: |
|
|
| | Model | Steps | Sample time | |
| | --- | --- | --- | |
| | BF16 | 8 | ~14 s | |
| | W4A8 (this repo) | 8 | ~7 s | |
| | int8_convrot (official) | 8 | ~6 s | |
| |
| Image quality is visually identical between BF16, W4A8 and the official |
| int8_convrot checkpoint. |
|
|
| ## Quantization format |
|
|
| Per quantized Linear layer the file stores: |
|
|
| - `<key>.weight` — int8, ConvRot-rotated packed int4 `[N, K/2]` |
| - `<key>.weight_s_rel` — fp8 e4m3fn group scale `[N, K/group_size]` |
| - `<key>.weight_s_channel` — fp32 channel scale `[N]` |
| - `<key>.weight_codebook` — fp32 Lloyd-Max codebook `[16]` |
| - `<key>.comfy_quant` — uint8 JSON `{"format": "asym_w4a8_int8", "group_size": 16, "convrot_groupsize": ...}` |
|
|
| 1D norms, biases, the embedding table and `cap_embedder.1` are kept in BF16. |
|
|
