QWEN Image Edit model which is designed to replace many different tools we’ve used like Inpainting, outpainting, controlnet, LoRA and its newest yet most powerful feature in context image editing with simple text prompts. No need for masks or segmentation, just ask the model to make the changes and most of the time it does just that.
Character consistency which had been an challenge is also addressed by this model just like it was with Flux Kontext in my earlier post.
QWEN Image Edit can generate consistently the provided subject in varying styles, environments and orientations. This means that we can now almost flawlessly create new images of our subject to build our clean and varied dataset.
In order to do this you first need to spend sometime figuring out how QWEN Image Edit actually alters and recreates a supplied image. Once you have this figured you can think of multiple prompts that will generate the person (man/woman) with the same consistent face again and again but in different clothes, expressions and posed angles. I have made some progress on this so this should give you a heads up in getting started.
Setup your ComfyUI Environment
Always make sure your ComfyUI is up to date, this is important when wanting to try newly released models or nodes. Next you need to download the model from the ComfyUI site. There are two models available now..the original Edit Model but then later a newer more improved model 2509 was also released.
- Qwen Image Edit Model – FP8 Model (qwen_image_edit_fp8_e4m3fn.safetensors) for low VRAM GPU (20.4gb) or get the full model qwen_image_edit_bf16.safetensors (40.9gb), place these in the folder
ComfyUI/models/diffusion_models/ - Qwen Image Edit 2509 Model – Newer FP8 Model (qwen_image_edit_2509_fp8_e4m3fn.safetensors) for low VRAM GPU (20.4gb) or get the full model qwen_image_edit_2509_bf16.safetensors (40.9gb), place these in the folder
ComfyUI/models/diffusion_models/ - Clip and VAE are the available here but you only need them if you haven’t already configured Qwen, if you are already using the model Qwen Image Edit uses the same.
If you want to run the ~40GB model then you want to use RTX5090 with 32 GB VRAM or bigger GPU, I can get cheap access to this GPU for less than $1 per hour at RunPod. Now if ComfyUI is running, make sure you restart it and you have Updated it!
Default workflow is now available via the Templates menu. Search for Qwen and you will find it.
My modified version of the workflow is also available below.
Generating LoRA Dataset
Once its all setup and ready you can get started with the generation. I have provided my prompts that I used for the below samples which you can download for free.
Here we have a sample image that we may want to use to generate the rest of the dataset for LoRA training.
Next you can install my Custom Nodes that are provided in my workflow (download below). Or search ComfyUI Manager for “WWAA Custom Nodes”.
Download Workflow
My prompt file contains 50 different variations (gender neutral and female version) that will help you generate the different sample but you can run it as much as you need to generate the sample image. Result is that we have below sample images that are generated with the consistent face however in different poses, clothes and backgrounds.
Resulting images can then be used in Ai Toolkit to train a Wan/Flux/Qwen LoRA or SDXL LoRA. I have many posts and videos on this topic so it can certainly help you further in LoRA training.
Thanks for stopping buy hope you found this useful. Let me know if you have any questions or comments.
You can also checkout my video where I walk you through this process using RunPod.
You may also like:
If you'd like to support our site please consider buying us a Ko-fi, grab a product or subscribe. Need a faster GPU, get access to fastest GPUs for less than $1 per hour with RunPod.io