Create and edit images with AI on your computer without paying anything
Qwen Image 2.1 is Alibaba's image AI, compressed to fit in 4.4 GB. It creates images from text and edits your own, with no account, no credits and no filters.
Steps
- Install ComfyUI, which is free: comfy.org.
- Download the model from Hugging Face: chatrlabs/Qwen-Image-2.1-Uncensored-Omni-Quantized. The 4.4 GB one is the file qwen-image-2.1-step90000-Q4_K_M.gguf.
- Put the file in the diffusion_models folder, inside ComfyUI's models folder.
- Download the text encoder and the VAE too, which don't come in that file. They are in Comfy-Org/Qwen-Image-2.1 and go in models/text_encoders and models/vae.
- In ComfyUI, load the model with the Unet Loader (GGUF) node. The same file works for creating images and for editing them.
What Qwen Image 2.1 is
It's Alibaba's image model. Its official repository, Qwen/Qwen-Image-2.1, has been on Hugging Face since September 14, 2026. With 7 billion parameters in the part that generates the image, it creates images from text and edits your own with the same model, outputs images with a transparent background and takes up to 10 reference images.
The version in this article was published by ChatR Labs on November 3, 2026. It removes the original model's filters and compresses it into several sizes so it fits on a home graphics card.
Specs
| Spec | Value |
|---|---|
| Original model | Qwen Image 2.1, by Alibaba (September 2026) |
| This version | ChatR Labs, November 3, 2026 |
| Size | From 4.2 to 14.2 GB, depending on the version |
| Graphics card | About 8 GB |
| What it does | Creates images from text and edits your own |
| Where it runs | ComfyUI and Diffusers |
| Downloads | 268,000 in the last month on Hugging Face, as of November 19, 2026 |
| License | Qwen Research, personal use only |
Which version to download
The bigger the file, the closer the result is to the uncompressed model. All of these work in ComfyUI.
| File | Size |
|---|---|
| Q4_0 | 4.20 GB |
| Q4_K_M | 4.47 GB |
| Q5_K_M | 5.21 GB |
| Q6_K | 6.00 GB |
| Q8_0 | 7.69 GB |
| BF16 | 14.23 GB |
Recommended settings
| Setting | Value |
|---|---|
| Steps | 35 |
| CFG | 5.0 |
| Sampler | er_sde, with the simple scheduler |
| Resolution | 768 x 1152 or 1024 x 1024 |
| Negative prompt | A single space |
These are the settings ChatR Labs recommends on the model card. The negative prompt can't be empty: if you leave it blank, CFG stops working.
It's for personal use: the license doesn't let you sell what you make.