7 min read ·
Qwen Image Edit in ComfyUI: workflow, fp8 and VRAM
How to run Qwen Image Edit in ComfyUI: the files you need, fp8 or bf16, how much VRAM it takes, Lightning speed-up LoRA files and Rapid AIO builds on a rented GPU.
Qwen Image Edit is an open model from the Qwen team (Alibaba) that edits images from a text instruction: change the background, clothes, lettering or style, or combine objects. The model is large, so the easiest way to run it is ComfyUI on a card with 24 GB of video memory or more. If you do not own such a card, rent one by the minute on gpu.nz.
What you need to run it
In ComfyUI the model is assembled from several files. The ComfyUI team publishes ready repackaged versions on Hugging Face; look for the Comfy-Org repository with Qwen Image Edit.
| File | ComfyUI folder | Purpose |
|---|---|---|
| The Qwen Image Edit model | models/diffusion_models | Image generation and editing |
| A text encoder based on Qwen2.5-VL | models/text_encoders | Reads the instruction and the picture itself |
| VAE | models/vae | Encodes and decodes the image |
| A speed-up LoRA (optional) | models/loras | Fewer generation steps, a faster result |
fp8 or bf16: how much video memory you need
The model has about 20 billion parameters. In full bf16 precision the weights take around 40 GB, in fp8 half of that. Memory guidelines:
| Variant | VRAM | Comment |
|---|---|---|
| fp8 | 24 GB | The main option for an RTX 3090 or 4090. ComfyUI offloads part of the weights to system memory |
| fp8 on 32 GB | 32 GB | RTX 5090: less offloading and more speed |
| bf16 | 48 GB and more | Full precision on L40S, A100 or H100 class cards |
These are estimates from file sizes, not measurements. Real use depends on the picture's resolution and the number of input images. If memory runs out, lower the resolution or switch to fp8.
Workflow: how to build it
- Update ComfyUI to a recent version: older ones do not know the nodes. ComfyUI Manager installs what is missing.
- In ComfyUI open the workflow templates (Browse Templates) and find the Qwen Image Edit example: a ready graph with the right nodes.
- Select the downloaded files in the nodes: the model, the text encoder and the VAE.
- Load the source picture, type an instruction («replace the background with mountains», «remove the text») and run.
- For iterations, use the result as the new input image and keep refining.
Lightning LoRA and Rapid AIO: speeding it up
- Lightning LoRA files are distilled speed-up weights that plug in like a regular LoRA and let you use few steps (the version name states how many, usually 4 or 8). Quality is slightly lower, but the result arrives much sooner.
- Rapid AIO builds are community merges that put the model and speed-up parts into a single file, which keeps the workflow simple. They are unofficial and quality depends on the author, so test on your own tasks.
- Time per picture depends on the card, resolution and steps. We do not publish numbers we have not measured: time one generation and multiply by the hourly price divided by 3600.
Running it on a rented GPU
Launch a server with the ComfyUI template on a card with 24 GB or more, for example an RTX 4090, download the files into the folders with wget and open the interface from the button on the server page. If you want a ready chat for photo edits with no graphs, use the Uncensored Image Chat template. For a language model from the same team, use the Qwen 3.8 template. Memory rules for other tasks are in the FLUX guide.
Questions and answers
- How much VRAM does Qwen Image Edit need?
- For fp8, plan on 24 GB with part of the weights offloaded to system memory; bf16 needs 48 GB or more.
- Which is better, fp8 or bf16?
- bf16 is more precise but twice as heavy. fp8 takes half the memory and usually gives a close result, so it is the default.
- What is a Lightning LoRA?
- Speed-up weights that plug in like a LoRA and let you generate in 4–8 steps instead of dozens. Quality is slightly lower, speed is higher.
- Where do the model files go in ComfyUI?
- The main model in models/diffusion_models, the text encoder in models/text_encoders, the VAE in models/vae and LoRA files in models/loras.
- Can I run Qwen Image Edit without my own graphics card?
- Yes: rent a ComfyUI server on a card with 24 GB or more and pay by the minute.