Custom workflow for the Clothes Try On LoRA by kingroka.
You have the following input options to choose from:
Enable Outfit Extraction : extracts the outfit from your input image (slot 1) and puts it on a white background;
Enable Outfit Transfer : transfers the outfit already extracted you provide (slot 2) to the character (slot 3);
Both : extracts the outfit from your input image (slot 1) and transfers it to the character (slot 3).
With version 1.2 you now also have the following upscaling options to choose from:
Enable Upscale Extraction : upscales the outfit extracted
Enable Upscale Transfer : upscales the final result
Qwen-Edit being limited to a size of 1 megapixels means you lose a lot of the fine details during the outfit transfer. That is a built-in limitation of the model and the LoRA that no workflow can ever fix, unfortunately. However, upscaling can hopefully restore some of the details. I'm using Real ESRGAN 2xPlus but feel free to experiment with other models (kingroka recommends SeedVR2, but this one requires a LOT of memory so be warned).
With the upscaler, you may want to play with the Denoise Ratio. For realistic images, something between 0.2 and 0.3 should be the sweet spot. For anything digital, you should probably put it back down to 0.
It may seem a bit conterintuitive but I recommend using the regular Qwen-Image Lightning LoRA that gives significantly better results than the Qwen-Image-Edit Lightning (and is the one used by kingroka in his own workflow). Once again, feel free to experiment to find what works best for you and the style you want to achieve.
I'm still trying to figure out how to automatically crop the outfit from the final result. On top of making it instantly usable, it should make the upscaler more efficient as it wouldn't have to deal with the outfit on the left of the image.
This workflow uses quite a few custom nodes so I recommend using ComfyUI Manager to get them all easily.
Description
You now have the following upscaling options :
Enable Upscale Extraction : upscales the outfit extracted
Enable Upscale Transfer : upscales the final result
FAQ
Comments (6)
Why MXToolkit with its random sigma and steps settings when I can't see any noticeable difference in output quality. Why not stick to a straightforward, standard workflow instead? It also doesn't perform well in full FP16 precision, leading to imprecise application (like changing face). Plus, the final result is still just two stitched images in one frame. I get it's not your fault. qwen edit doesn't support direct reference images, but thanks for the visually slick workflow, I guess?
I haven't tried your new scheme yet, but I also use the Qwen-Image Lightning LoRA for 8 steps, the second version has recently been released, but I haven't installed it yet. I also tried SeedVR2, but it's unnecessarily long and heavy. As for how to automatically separate clothes, it is done very simply. You first need to automatically combine both images into one, for this they must be the same size, which is also done automatically, and then split into two (a / 2 where a is the width of the image), and then the gray edges need to be trimmed with a mask. That's what I did.
And I also noticed that the face changes depending on what kind of clothes you use for carrying. For example, if you take clothes from a fantasy character and put them on a real person. There is a high probability that the face will change too. Or if the clothes are too big in the head area (hat, scarf, collar, etc.). In this case, there may also be problems with the face. It is better to wear clothes that are worn by a real person, rather than just a suit on a white background.
How can I change the output size without using the upscaler? I can't find what param sets the img size for the extracted results (and the stitched result) Great wf though!
I tried your new workflow, but the face still changes slightly. To crop the left part of the image before upscaling, you either need to calculate the percentage ratio of the images - pixel > % > pixel = cropping. Or initially stitch the image using a different node where these proportions will be equal, and then simply divide by 2. -- an example of the second option https://civitai.com/posts/22607815
workflow too convoluted.