# Qwen-Image GGUF Guide to ComfyUI Setup and Quantization > Published 2026-04-03, updated 2026-09-05 ยท https://www.promptzone.com/divya_watanabe/qwen-image-gguf-lightweight-ai-for-image-generation-12ln --- title: Qwen-Image GGUF Guide to ComfyUI Setup and Quantization published: true description: Load Qwen-Image GGUF in ComfyUI with the correct custom loader, encoder, and VAE, then evaluate quantization tradeoffs. tags: ai,comfyui,imagegeneration --- To use Qwen-Image GGUF in ComfyUI, install city96's ComfyUI-GGUF loader and load a converted image model alongside the required text encoder and VAE. The conversion card supplies downloads, folder locations, and an example workflow. These files represent Alibaba Qwen's 20B image model in GGUF format. [Conversion card][gguf] [Qwen repository][repo] ## What are the key facts about Qwen-Image GGUF? | Field | Verified information | | --- | --- | | Developer | Alibaba Qwen develops the original model; city96 publishes this GGUF conversion. [Qwen repository][repo] [Conversion card][gguf] | | Released | Base weights: August 4, 2025. A conversion release date is not published in the cited GGUF model card. [Base card][base] [Conversion card][gguf] | | Type | Quantized text-to-image model distribution for a GGUF-compatible loader. [Conversion card][gguf] | | Size or parameters | The original image model is 20B; the conversion card identifies it as a direct conversion, not a separate architecture. [Qwen repository][repo] [Conversion card][gguf] | | License and access | Apache 2.0; downloadable GGUF files retain the original model's restrictions and license terms. [Conversion card][gguf] | | Where it runs | ComfyUI with the ComfyUI-GGUF custom node and the required encoder and VAE. [Conversion card][gguf] | Start with the conversion publisher's model card, because it names the associated components and links an example graph. A generic GGUF tutorial may describe a different image model or a language-model workflow. ## How does Qwen-Image GGUF quantization work? ComfyUI-GGUF provides support for storing compatible diffusion-transformer weights with lower-bit quantization. Its purpose is to make another weight representation available within ComfyUI, where users can test memory and quality tradeoffs using their existing image-generation workflow. [Loader README][ggufnode] The Qwen conversion card describes a specific treatment for several quantizations, including Q5_K_M, Q4_K_M, and lower-bit variants: the first and last layers are retained at higher precision. This is a detail of the published conversion method, so do not assume that every layer uses a single uniform precision from the filename. [Conversion card][gguf] The publisher supplies a Qwen-specific workflow with a GGUF image loader, Qwen text-encoder selection, and the Qwen VAE. That gives you a concrete starting graph to inspect before experimenting with different conversion files. [Example workflow][ggufworkflow] A useful evaluation includes a simple illustration, a face, a textured object, and a short text-bearing image. Choose examples relevant to your work and inspect the same features across candidate quantizations. These are suggested test cases, not a claim that any particular quantization has passed them. ## What are the limits of Qwen-Image GGUF in ComfyUI? The conversion card labels its displayed output as a single sample that is not necessarily representative. It does not publish a universal GPU-memory minimum or a fixed image-generation time. Treat the sample as evidence that the illustrated workflow exists, then measure the workload you intend to run. [Conversion card][gguf] Downloading the GGUF transformer alone does not complete the pipeline. The model card separately requires a Qwen2.5-VL-7B text encoder and Qwen-Image VAE. A smaller image-model file therefore does not by itself establish the memory use of the full graph. [Conversion card][gguf] The generic ComfyUI-GGUF README calls the project a work in progress and describes LoRA loading as experimental. If an adapter is part of your intended workflow, verify that combination separately after establishing a working baseline. [Loader README][ggufnode] The Qwen conversion card places its image model in `models/diffusion_models`; the generic ComfyUI-GGUF README uses `models/unet`. Follow the Qwen-specific layout for the supplied graph and select the downloaded file in its loader. [Conversion card][gguf] [Loader README][ggufnode] ## How do you install and load Qwen-Image GGUF in ComfyUI? Update ComfyUI, then install ComfyUI-GGUF in the Python environment that actually runs the application. From the ComfyUI installation directory, the normal Git and Python installation can be written as follows, adapting the folder placement in the loader's README. [Loader README][ggufnode] ```bash git clone https://github.com/city96/ComfyUI-GGUF custom_nodes/ComfyUI-GGUF python -m pip install --upgrade gguf ``` For a Windows portable installation, use the bundled Python executable and the portable commands in the official loader README. Installing a dependency into a different system Python does not configure the environment named by those commands. [Loader README][ggufnode] Download one GGUF image-model file from the city96 Qwen repository. Put it in `ComfyUI/models/diffusion_models`, the Qwen encoder in `ComfyUI/models/text_encoders`, and `qwen_image_vae.safetensors` in `ComfyUI/models/vae`. The conversion card links both standard and GGUF text-encoder options. [Conversion card][gguf] Import the example workflow linked from the card. Its `UnetLoaderGGUF` selects the image model, while its encoder loaders specify `qwen_image` as the model type. The graph includes both ordinary and GGUF encoder-loading options; inspect which path is connected and select the actual file you downloaded. [Example workflow][ggufworkflow] Use the graph's settings for an initial run before adding LoRAs or replacing other nodes. Save the workflow and the exact conversion filename with the output. Then change only the image-model quantization and compare spelling, outlines, textures, runtime, and memory behavior against the baseline. If you need help identifying the original components, see the sibling [native Qwen-Image ComfyUI guide](/anika_bose/qwen-image-powerful-ai-art-tool-for-comfyui-lna). The sibling [GPU planning guide](/harper_korhonen/qwen-25-vision-powering-ai-with-gpu-specs-2fp1) explains why encoder, transformer, and VAE residency should be considered separately. ## How does Qwen-Image GGUF compare with native and Nunchaku setups? | Route | Model representation and loader | Suggested decision criterion | | --- | --- | --- | | Native ComfyUI Qwen-Image | Documented Safetensors image checkpoint and native loaders. [Guide][comfy] | Use as the reference setup for your project. | | city96 Qwen-Image GGUF | Converted GGUF checkpoint and ComfyUI-GGUF image loader. [Card][gguf] | Evaluate the available quantizations on your workload. | | Nunchaku Qwen-Image | SVDQuant weights and the Nunchaku runtime. [Card][nun] | Check a specialized runtime's GPU and package requirements. | These routes require their own documented loaders. The [ComfyUI complete guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) provides the broader graph concepts needed to compare them without losing track of which components changed. ## What should you check before using Qwen-Image GGUF? ### What does Qwen-Image GGUF quantize? The city96 Qwen-Image GGUF repository provides a direct conversion of Qwen's image model, whose underlying architecture is 20B. The conversion changes weight representation, and the pipeline still needs a separately loaded text encoder and VAE. [Conversion card][gguf] [Qwen repository][repo] ### Which ComfyUI loader supports Qwen-Image GGUF? The published Qwen-Image GGUF workflow uses ComfyUI-GGUF's `UnetLoaderGGUF` for the image transformer. Install that integration, import the supplied graph, and select the converted file in the GGUF loader. [Conversion card][gguf] [Example workflow][ggufworkflow] ### Does Qwen-Image GGUF need a text encoder and VAE? Qwen-Image GGUF requires a Qwen2.5-VL-7B text encoder and Qwen-Image VAE alongside the converted image model. The city96 model card links standard and GGUF encoder options and provides a separate VAE download. [Conversion card][gguf] ### Which Qwen-Image GGUF quantization should I choose? Compare Qwen-Image GGUF files on the prompts, image dimensions, and hardware you intend to use. The publisher labels its displayed output as a single sample, so evaluate spelling, detail, speed, and complete-pipeline memory use before choosing a quantization. [Conversion card][gguf] ## Sources - [city96 Qwen-Image GGUF model card][gguf] - [ComfyUI-GGUF installation and usage][ggufnode] - [city96 Qwen-Image GGUF example workflow][ggufworkflow] - [Qwen-Image model card][base] - [Qwen-Image official repository and release history][repo] - [ComfyUI official Qwen-Image workflow guide][comfy] - [Nunchaku Qwen-Image model card][nun] [gguf]: https://huggingface.co/city96/Qwen-Image-gguf [ggufnode]: https://github.com/city96/ComfyUI-GGUF [ggufworkflow]: https://huggingface.co/city96/Qwen-Image-gguf/resolve/main/media/qwen-image_workflow.json [base]: https://huggingface.co/Qwen/Qwen-Image [repo]: https://github.com/QwenLM/Qwen-Image [comfy]: https://docs.comfy.org/tutorials/image/qwen/qwen-image [nun]: https://huggingface.co/nunchaku-tech/nunchaku-qwen-image ## Related guides on PromptZone - [Local LLMs 2026: Run Llama, Mistral, Qwen on Your Hardware](/lukas_tanaka/local-llms-2026-run-llama-mistral-qwen-on-your-hardware-complete-guide-32k)