# How to Run Qwen-Image-2512 Locally in ComfyUI **Qwen-Image-2512** is the December update to Qwen's text-to-image foundational models. The model is the top performing open-source diffusion model and this guide will teach you how to run it locally via [Unsloth](https://github.com/unslothai/unsloth) GGUF and ComfyUI. Qwen-Image-2512 features: more realistic looking people; richer details in landscapes/textures; and more accurate text rendering. **Uploads:** [GGUF](https://huggingface.co/unsloth/Qwen-Image-2512-GGUF) • [FP8](https://huggingface.co/unsloth/Qwen-Image-2512-FP8) • [4-bit BnB](https://huggingface.co/unsloth/Qwen-Image-2512-unsloth-bnb-4bit) The quants use [Unsloth Dynamic](https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs) methodology which upcasts important layers to higher precision to recover more accuracy. Thank you Qwen for allowing Unsloth day 0 support. ## 📖 ComfyUI Tutorial To run, you don't need a GPU, just a CPU with RAM will work. For best results, ensure your total usable memory (RAM + VRAM / unified) is larger than the GGUF size; e.g. 4-bit (Q4\_K\_M) `unsloth/Qwen-Image-Edit-2512-GGUF` is 13.1 GB, so you should have 13.2+ GB of combined memory. [ComfyUI](https://github.com/Comfy-Org/ComfyUI) is an open-source diffusion model GUI, API, and backend that uses a node-based (graph/flowchart) interface. This guide will focus on machines with CUDA, but instructions to build with on Apple or CPU are similar. ### #1. Install & Setup To install ComfyUI, you can download the desktop app on Windows or Mac devices [here](https://www.comfy.org/download). Otherwise, to setup ComfyUI for running GGUF models run the following: ```bash mkdir comfy_ggufs cd comfy_ggufs python -m venv .venv source .venv/bin/activate git clone https://github.com/Comfy-Org/ComfyUI.git cd ComfyUI pip install -r requirements.txt cd custom_nodes git clone https://github.com/city96/ComfyUI-GGUF cd ComfyUI-GGUF pip install -r requirements.txt cd ../.. ``` ### #2. Download Models Diffusion models typically need 3 models. A Variational AutoEncoder (VAE) that encodes image pixel space to latent space, a text encoder to translate text to input embeddings, and the actual diffusion transformer. You can find all Unsloth diffusion GGUFs in our [Collection here](https://huggingface.co/collections/unsloth/unsloth-diffusion-ggufs). Both the diffusion model and text encoder can be GGUF format while we typically use safetensors for the vae. According to [Qwen's repo](https://huggingface.co/Qwen/Qwen-Image-2512/blob/main/text_encoder/config.json), we shall use Qwen2.5-VL and not [Qwen3-VL](https://unsloth.ai/docs/models/tutorials/qwen3-how-to-run-and-fine-tune/qwen3-vl-how-to-run-and-fine-tune). Let's download the models we will use (Note: You can also also use our [FP8 upload](https://huggingface.co/unsloth/Qwen-Image-2512-FP8) in ComfyUI): ```bash cd models ## Diffusion Models curl -L -C - -o unet/qwen-image-2512-Q4_K_M.gguf \ https://huggingface.co/unsloth/Qwen-Image-2512-GGUF/resolve/main/qwen-image-2512-Q4_K_M.gguf curl -L -C - -o unet/qwen-image-edit-2511-Q4_K_M.gguf \ https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/resolve/main/qwen-image-edit-2511-Q4_K_M.gguf ## Text Encoder + Vision Tower + VAE curl -L -C - -o text_encoders/Qwen2.5-VL-7B-Instruct-UD-Q4_K_XL.gguf \ https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/Qwen2.5-VL-7B-Instruct-UD-Q4_K_XL.gguf curl -L -C - -o text_encoders/Qwen2.5-VL-7B-Instruct-mmproj-BF16.gguf \ https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/mmproj-BF16.gguf curl -L -C - -o vae/qwen_image_vae.safetensors \ https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors ``` See GGUF uploads for: [Qwen-Image-2512](https://huggingface.co/unsloth/Qwen-Image-2512-GGUF), [Qwen-Image-Edit-2511](https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF), and [Qwen-Image-Layered](https://huggingface.co/unsloth/Qwen-Image-Layered-GGUF) {% hint style="warning" %} The format of the vae and diffusion model might be different than the diffusers checkpoints if using checkpoints other than the ones above. Only use checkpoints that are compatible with ComfyUI. {% endhint %} These files must be in the correct folders for ComfyUI to see them. In addition the vision tower stored in the mmproj file must use the same prefix as the text encoder. Download reference images to be used later as well: ```bash curl -L -C - -o ../input/sloth1.jpg \ "https://unsloth.ai/cgi/image/_1d5a5685-2d88-44ca-b50f-ba432cd646ef_9CGCY8lvw4D9JkOdueqsk.jpeg?width=1920&quality=80&format=jpeg" curl -L -C - -o ../input/sloth2.jpg \ "https://unsloth.ai/cgi/image/UnSloth_GPU_Front_-_Confetti_ArcSk-MR4MMN215UutOFZ.png?width=1920&quality=80&format=jpeg" ``` ### #3. Workflow and Hyperparameters For more info you can also view our detailed [#workflow-and-hyperparameters-1](https://unsloth.ai/docs/blog/comfyui#workflow-and-hyperparameters-1 "mention") Guide. Navigate to the main ComfyUI directory and run: ```bash python main.py ``` {% hint style="info" %} `python main.py --cpu` to run with CPU, but will be slow. {% endhint %} This will launch a web server that allows you to access `https://127.0.0.1:8188` . If you are running this on the cloud, you'll need to make sure port forwarding is setup to access on your local machine. Workflows are saved as JSON files embedded in output images (PNG metadata) or as separate `.json` files. You can: * Drag & drop an image into ComfyUI to load its workflow * Export/import workflows via the menu * Share workflows as JSON files Below are two examples of Qwen-Image-2512 and Qwen-Image-Edit-2511 json files which you can download and use: {% file src="" %} For our workflow, we default to **1024×1024** as a practical middle ground. While the model supports native resolution (1328×1328), generating at native typically increases runtime by **\~50%**. Since GGUF adds overhead and 40 steps is a relatively long run, 1024×1024 keeps generation time reasonable. If needed, you can increase resolution to 1328. {% hint style="warning" %} For more realistic results, skip keywords like “photorealistic” or “digital rendering” or “3d render” and use terms like “photograph” instead. {% endhint %} {% hint style="info" %} For negative prompts, it’s best to use an NLP-style approach: describe in **natural language** what you *don’t* want in the image. Packing in too many keywords can hurt results instead of making it more specific. {% endhint %} {% file src="" %} {% columns %} {% column %} Instead of setting up the workflow from scratch you can download the workflow here. Load it into the browser page by clicking the Comfy Logo -> File -> Open -> Then choose the `unsloth_qwen_image_2512.json` file you just downloaded. It should look like the below: {% endcolumn %} {% column %}
{% endcolumn %} {% endcolumns %}
This workflow is based on the official ComfyUI published workflow except it uses the GGUF loader extension, and is simplified to illustrate text to image functionality. ### #4. Inference ComfyUI is highly customizable. You can mix models and create extremely complex pipelines. For a basic text to image setup we need to load the model, specify prompt and image details, and decide on a sampling strategy. #### **Upload Models + Set Prompt** We already downloaded the models, so we just need to pick the correct ones. For Unet Loader pick `qwen-image-2512-Q4_K_M.gguf`, for CLIPLoader pick `Qwen2.5-VL-7B-Instruct-UD-Q4_K_XL.gguf`, and for Load VAE pick `qwen_image_vae.safetensors`. {% hint style="info" %} For more realistic results, skip keywords like “photorealistic” or “digital rendering” or “3d render” and use terms like “photograph” instead. {% endhint %} You can set any prompt you'd like, and also specify a negative prompt. The negative prompt helps by telling the model where to steer away from. {% hint style="info" %} For negative prompts, it’s best to use an NLP-style approach: describe in **natural language** what you *don’t* want in the image. Packing in too many keywords can hurt results instead of making it more specific. {% endhint %} #### **Image Size + Sampler Parameters** The Qwen Image model series supports different image sizes. You can make rectangular shapes by setting the values of width and height. For sampler parameters, you can experiment with different samplers other than euler, and more or less sampling steps. The workflow has steps set to 40, but for quick tests 20 might be good enough. Change the `control after generate` setting from randomize to fixed if you want to see how different settings change outputs. #### **Run** Click Run and an image will be generated in about 1 minute (30 seconds for 20 steps). That output image can be saved. The interesting part is that the metadata for the entire comfy workflow is saved in the image. You can share and anyone can see how it was created by loading it in the UI.
{% hint style="info" %} If you're encountering blurry/bad images, raise shift to 12-13! solves most issues with bad outputs. {% endhint %} #### **Multi Reference Generation** A key feature of Qwen-Image-Edit-2511 is multi reference generation where you can supply multiple images to use to help control generation. This time load the `unsloth_qwen_image_edit_2511.json`. We will use most of the same models but switching `qwen-image-2512-Q4_K_M.gguf` to `qwen-image-edit-2511-Q4_K_M.gguf` for the unet. The other difference this time are extra nodes to select images to reference, which we've downloaded earlier. You'll notice the prompt refers to both `image 1` and `image 2` which are prompt anchors for the images. Once loaded click Run, and you'll see an output that creates our two unique sloth characters together while preserving their likeness.

Final result made from images to the right:

## 🤗 D**iffusers Tutorial** We have also uploaded a [Dynamic 4-bit BitsandBytes](https://huggingface.co/unsloth/Qwen-Image-2512-unsloth-bnb-4bit) quantized version which can be run in Hugging Face's `diffusers` library. Once again, it uses Unsloth Dynamic where important layers are upcasted to higher precision. Run `Qwen-Image-2512-unsloth-bnb-4bit` with the code below: ```python from diffusers import DiffusionPipeline import torch pipe = DiffusionPipeline.from_pretrained( "unsloth/Qwen-Image-2512-unsloth-bnb-4bit", torch_dtype=torch.bfloat16, ).to('cuda') # uncomment if you run out of memory # pipe.enable_model_cpu_offload() output = pipe( prompt="a kawaii sloth playing the drums", negative_prompt="blurry, unfocused", num_inference_steps=20, true_cfg_scale=4.0, ) # Save output image = output.images[0] image.save('sample.png') ``` ## 🎨 **stable-diffusion.cpp Tutorial** If you want to run the model in stable-diffusion.cpp, you can follow our [step-by-step guide here](https://unsloth.ai/docs/models/tutorials/qwen-image-2512/stable-diffusion.cpp).