Qwen-Image-2.1 GGUF Quants
This repository provides quantized GGUF checkpoints for Qwen/Qwen-Image-2.1, optimized for low-VRAM inference in ComfyUI using the ComfyUI-GGUF custom node.
Quantizing the diffusion transformer drastically reduces memory pressure during generation while preserving sharp detail, composition, and prompt alignment.
📦 Quantization Breakdown & File Details
| Filename | Quant Type | Size | Recommended VRAM / Profile |
|---|---|---|---|
qwen_image_2.1_Q8_0.gguf | Q8_0 | 7.59 GB | 12 GB+ (Near-lossless fidelity) |
qwen_image_2.1_Q6_K.gguf | Q6_K | 5.88 GB | 10–12 GB (High quality sweet spot) |
qwen_image_2.1_Q5_K_M.gguf | Q5_K_M | 5.01 GB | 8–10 GB (Balanced performance & memory) |
qwen_image_2.1_Q4_K_M.gguf | Q4_K_M | 4.19 GB | 6–8 GB (Standard consumer GPU baseline) |
qwen_image_2.1_Q4_K_S.gguf | Q4_K_S | 4.06 GB | 6–8 GB (Compact 4-bit) |
qwen_image_2.1_Q3_K_M.gguf | Q3_K_M | 3.19 GB | 4–6 GB (Extreme budget / minimum VRAM) |
🖼️ Sample Generations (Q4_K_M)
🧩 Required Components (Text Encoders & VAE)
To run these models in ComfyUI, you will need the matching text encoders and VAE:
-
Text Encoders:
- Download from: Comfy-Org/Qwen-Image-2.1 Text Encoders
- Place files into:
ComfyUI/models/clip/(orComfyUI/models/text_encoders/)
-
VAE:
- Download from: Comfy-Org/Qwen-Image-2.1 VAE
- Place files into:
ComfyUI/models/vae/
-
Diffusion Model (This Repo):
- Download your preferred
.gguffile from above. - Place into:
ComfyUI/models/diffusion_models/(orComfyUI/models/unet/)
- Download your preferred
🎨 Included Ready-to-Use Workflows
Both Text-to-Image and Image-to-Image editing workflows are packaged in this repo:
- Text-to-Image: Qwen_Image_2.1_GGUF_Text2Image.json
- Image-to-Image / Editing: Qwen_Image_2.1_GGUF_Image2Image_Edit.json
How to Use:
- Ensure you have installed ComfyUI-GGUF (search for
ComfyUI-GGUFinside the ComfyUI Manager). - Drag and drop either of the
.jsonworkflow files into your ComfyUI workspace. - In the Unet Loader (GGUF) node, select the downloaded
.gguffile. - Verify your CLIP/Text Encoder and VAE node loaders point to the files downloaded above.
- Queue prompt and generate!
👏 Acknowledgments & Credits
- Base model developed by the Qwen Team / Alibaba Cloud: Qwen/Qwen-Image-2.1.
- ComfyUI integration assets and pipeline weights provided by Comfy-Org.
- GGUF quantization powered by llama.cpp and ComfyUI-GGUF.