🛠️ Desaign Qwen Inpaint & Edit
A brand-new addition to the Desaign workflow suite — this one is powered by Qwen, focusing on flexible, high-quality inpainting and image editing with full creative freedom.
Unlike traditional inpainting setups, this workflow doesn’t just patch pixels — it seamlessly integrates characters, objects, or edits into your image while preserving its overall mood, colors, and resolution. Think of it as a next-gen editing desk: precise when you need control, fluid when you want to experiment.
🔑 Key Features
Qwen-powered inpainting → natural language editing and insertion with powerful uncensored outputs.
Dual modes → choose between Inpaint Edit (modify only the masked area) or Whole Image Edit (reimagine the entire composition).
Multi-image integration → work with up to 3 different images at once, easily merging characters or elements across them.
Resolution Master → default scaling adapts your input (resized to 1152px for optimal Qwen performance), with a toggle to preserve original resolution.
Fast group bypasser → enable/disable additional images and modules without breaking the workflow.
Uncensored freedom → Qwen keeps the door open for creative exploration, no matter your subject matter.
⚙️ How It Works
Load your base image.
Mask the area you want to edit: right-click → Open in Mask Editor.
Choose your mode:
Inpaint Edit → updates only the masked region.
Whole Image Edit → lets Qwen reinterpret the full image.
(Optional) Add up to 3 source images, then enable them via the fast group bypasser.
Set resolution options:
Default = resized to 1152px.
To keep custom sizing, disable USE IMAGE 1 RESOLUTION.
Prompt Qwen with your desired change (e.g., “replace this with a character holding a lantern in the rain”).
Generate & compare results.
⚠️ Notes
The quality depends on your mask size and the detail of your prompt. Smaller, well-defined zones generally produce cleaner edits.
Multi-image blending opens new creative doors — but keep lighting, perspective, and style consistent for best results.
For large edits, Whole Image Edit may yield more coherent results.
✨ In short: Desaign Qwen Inpaint & Edit is like having Nano Banana at home — but with the uncensored freedom you always wanted. It’s powerful, flexible, and built for creators who want precision and imagination in the same tool.
Description
FAQ
Comments (12)
Great, but can you record a video about how to use it, with best practices for each function?
or lora recommendations,...
it's honesthly deceptively easy to use, most control are buttons, and if you have a doubt there is a how to use guide in the notes.
we have a youtube channel in project for those stuff, but time is missing right now.
if you have any additional question feel free to join our discord in my profile page. i'll be glad to help ; )
how do I set resolution options. it seems to do what it wants. is there a way to get a fixed output resolution? or the same resolution as the input image?
For now the resolution is resized by 1152 on the longest size, if you activate the same resolution boolean node, it will use the resolution of the image 1. Otherwise you can set it manually via the resolution master. I plan to rework the workflow, so don't hesitate to tell us what you think is missing
That resize thing did not work for me. It kinda does what it wants but i never get the input resolution
@chrisbraeuer41172035Â try the beta version, if you have any trouble you can join the discord in my profile description, i'll be glad to help you
@Archi_teknik101Â i just saw this comment, and i have a suggestion for the work flow, why not add in the availability to use comfyui_controlnet_aux (OpenposePreprocessor and DepthAnythingV2Preprocessor ) where if you have selected "true" in a Boolean & have enabled the control net group nodes for image 2 or 3 , then it will activate the open pose control net if you have selected "false" & have enabled the control net node it will activate the depth control net, but if you don't enable the control net group it will just use the image in 2 or 3 image.
I suggest this because the model has built in capability to use open pose and depth maps. this would allow you to make those control net images to use.
@ByteCrafter controlnets are a good addition, i'll most defenetly add it, thanks for the suggestion !
@Archi_teknik101 another thing i have just seen https://www.reddit.com/r/StableDiffusion/comments/1o01e6i/totally_fixed_the_qwenimageedit2509_unzooming/ might be worth looking into, moving the VEA apparently it helps to keep pixel accuracy
TL;DR :
Disconnect the VAE input from the TextEncodeQwenImageEditPlus node
Add a VAE Encode per source, and chained ReferenceLatent nodes, one per source also.
...
Profit !
Long version :
Here is an example of pixel-perfect match between an edit and its source. First image is with the fixed workflow, second image with a default workflow, third image is the source. You can switch back between the 1st and 3rd images and see that they match perfectly, rendered at a native 1852x1440 size.
@ByteCrafter i fairly tested and reviewed it, not to be too harsh but i found it solves a problem that isn't even present in the model. i think the consistency no shift tab in the desaign workflow eliminate those reference latent and multi vae encode, it seems like a trick more than anything really. the chinese node is super powerful with the consistency lora, much more than any of my test with the workflow on reddit. thanks for the recommendation it was at least worth a try.
@Archi_teknik101Â i only suggested it as the consistancy no shift before, for me has caused some issues where its shifted the image quite a bit. just the same way the create new image one does that was all, and then using the oen linked i noticed it never did that, that was all.