To run Qwen-Image in ComfyUI, import the official text-to-image workflow and load its diffusion model, Qwen2.5-VL text encoder, and VAE. These components run Alibaba Qwen's 20B image model, whose weights are available under Apache 2.0. Select the three files in their loader nodes, enter a prompt, and queue the first image. ComfyUI guide Model card
What are the key facts about Qwen-Image in ComfyUI?
| Field | Verified information |
|---|---|
| Developer | Alibaba's Qwen team develops the model; ComfyUI documents its native workflow. ComfyUI guide |
| Released | August 4, 2025, for the original Qwen-Image weights. Model card |
| Type | Text-to-image generation using an MMDiT diffusion transformer. Repository |
| Size or parameters | 20 billion parameters for the image model; the workflow separately loads a Qwen2.5-VL text encoder. ComfyUI guide |
| License and access | Apache 2.0 model weights, available through Qwen and the ComfyUI guide's linked model downloads. Model card ComfyUI guide |
| Where it runs | Local ComfyUI installations; the official guide also links a Comfy Cloud template. ComfyUI guide |
Keep the image transformer, text encoder, and VAE filenames together in your setup notes. ComfyUI loads each component separately in this workflow. ComfyUI guide
What can you generate with the native Qwen-Image workflow?
Qwen emphasizes text within generated images, particularly English and Chinese lettering, together with a range of visual styles. The model card demonstrates signs, posters, photographic scenes, paintings, and illustrations. These examples make text-bearing compositions a sensible place to begin your own evaluation. Model card
ComfyUI's practical contribution is a documented graph with explicit file loaders and sampling controls. You can inspect which model files a saved workflow expects and compare a changed prompt against a known setup. Use that visibility to establish a baseline before adding more components. ComfyUI guide
For a first prompt, try a simple poster with a single title, a clear subject, and a specified background. For example: a blue ceramic teapot on a cream background, with the title “Afternoon Tea” above it. This is a suggested test prompt, not a reported model output or a benchmark.
Review spelling, composition, and the relationship between the requested objects. Once the basic graph works, build a small set of prompts representing the images you actually need. That set will be more useful for future configuration changes than a single attractive result.
What causes Qwen-Image setup problems in ComfyUI?
The ComfyUI guide lists the BF16 and FP8 image model files at 40.9 GB and 20.4 GB respectively. Those are download sizes for particular files, not complete runtime memory requirements; the same workflow also loads an encoder and VAE. ComfyUI guide
The guide publishes a hardware-specific timing example, but it does not establish one minimum VRAM capacity or one generation speed for every configuration. Avoid choosing hardware from the checkpoint filename alone. For memory planning, record the exact files, dimensions, and workflow settings you test. ComfyUI guide
Missing nodes can indicate an outdated ComfyUI installation or an import failure, according to the official tutorial. A graph that loads successfully can still have missing model files or incorrect loader selections, so check the startup log and the graph separately. ComfyUI guide
The native text-to-image setup is also a specific workflow. For instruction-based edits to an existing image, Qwen documents a separate Qwen-Image-Edit checkpoint and pipeline. Select the model and example that match the operation you want to perform. Editing model card
How do you install and run Qwen-Image in ComfyUI?
First update ComfyUI using its normal update procedure. Open the template library and search for Qwen-Image, or download the workflow linked in the official tutorial and drag it into the interface. The guide explains that an absent template may indicate an outdated installation. ComfyUI guide
Next download the three files specified for the original FP8 workflow. Keep the filenames intact so they remain easy to match against the loader nodes. The locations below come directly from ComfyUI's documented model layout. ComfyUI guide
ComfyUI/models/diffusion_models/qwen_image_fp8_e4m3fn.safetensors
ComfyUI/models/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
ComfyUI/models/vae/qwen_image_vae.safetensors
From a terminal opened in your ComfyUI installation directory, you can create the required folders before copying the downloads:
mkdir -p models/diffusion_models models/text_encoders models/vae
Set Load Diffusion Model to the image checkpoint, Load CLIP to the named text encoder, and Load VAE to the Qwen image VAE. Set the image dimensions in the workflow's EmptySD3LatentImage node and enter your prompt in the text-encoding node. These are the loader and input controls identified in the official walkthrough. ComfyUI guide
For the initial run, use the original-model configuration supplied by the workflow and keep acceleration disabled. Queue an image, inspect the result, and save a copy of the graph together with the prompt and model filenames. This suggested baseline makes later changes easier to evaluate.
When you want fewer sampling steps, use the separate Qwen-Image-Lightning setup guide. The official ComfyUI page includes a Lightning branch, but checkpoint and adapter pairing deserve their own check rather than being treated as an automatic speed toggle. ComfyUI guide
How does native Qwen-Image compare with GGUF and Edit?
| Route | What changes | Suggested use |
|---|---|---|
| Native Qwen-Image workflow | Uses the documented ComfyUI loaders and split model files. Guide | Establish a text-to-image baseline. |
| Qwen-Image GGUF | Uses a converted image model and the ComfyUI-GGUF custom loader. Conversion card | Evaluate a different quantization setup. |
| Qwen-Image-Edit | Uses an editing checkpoint and input image with instructions. Card | Modify existing visual material. |
Choose the route based on the task and the supported loader, then compare results using your own prompts. The ComfyUI complete guide covers the broader interface and workflow concepts used throughout this setup.
What should you check when setting up Qwen-Image in ComfyUI?
Can Qwen-Image run outside ComfyUI?
Qwen-Image has downloadable weights and an official Diffusers example for Python inference. ComfyUI offers a separate native workflow for generating images through a visual graph. Model card ComfyUI guide
Which files does Qwen-Image need in ComfyUI?
The original FP8 Qwen-Image workflow uses qwen_image_fp8_e4m3fn.safetensors, qwen_2.5_vl_7b_fp8_scaled.safetensors, and qwen_image_vae.safetensors. Place them in ComfyUI/models/diffusion_models, ComfyUI/models/text_encoders, and ComfyUI/models/vae, respectively, then select each file in its loader. ComfyUI guide
Which text encoder does Qwen-Image use in ComfyUI?
The documented Qwen-Image ComfyUI workflow uses a Qwen2.5-VL-7B text encoder. It loads separately from the 20B image transformer and the Qwen image VAE. ComfyUI guide
How do I fix missing Qwen-Image nodes in ComfyUI?
For missing nodes in the native Qwen-Image workflow, update ComfyUI and inspect startup messages for failed imports. ComfyUI notes that newly documented core nodes may require a nightly build until the next stable release. ComfyUI guide
Sources
- Qwen-Image model card
- Qwen-Image official repository and release history
- ComfyUI official Qwen-Image workflow guide
- Qwen-Image-Edit model card
- city96 Qwen-Image GGUF model card
Top comments (0)