V20 Model Feedback: Concerning Regressions in Skin Texture and Fabric Detail
I'll add my two cents. This is my second day testing the NSFW v20 model, and in my opinion, it has gotten worse compared to version 19. The characters' skin has become more "waxy," and artifacts have started appearing on the body in the form of random imitations of moles. However, the number of these "moles" is just insane and looks more like debris stuck to wax than natural human skin. Yes, version 19 also suffered from this, but not on such a crazy scale, and to achieve naturalness, it was enough to do a refiner pass using a photorealistic SDXL model with a denoise strength of 0.09-0.13. In version 20, this no longer works: you need to increase the denoise value to 0.3-0.4, and in that case, all the advantages of Qwen Edit are simply negated. Character consistency is completely lost, the environment turns from realistic into some kind of surreal mess, more like a frame from a dream. In short, welcome back to all the SD flaws that were so annoying...
Separately about fabric textures: in version 20, they are gone. There is something that's hard to even call fabric. If we just edit a picture "as is," for example, correcting the lighting, then everything is almost okay, within the capabilities of an 8-step LoRA. But as soon as you try to change the pose or angle—that's it, the fabric turns into something I can't even describe in words :)
P.S. A special thanks for the node. It saved my frayed nerves. I hesitated to install it for a long time, but when I finally did, the generations became absolutely marvelous! It's wonderful.
P.P.S. Please don't throw slippers at me for mistakes in the text: the translation to English was done with the help of AI :)
increase steps to 6 instead of 4, Skin Texture and Fabric Detail will be greatly improved
My first impression after using version v20 was that the consistency of the characters' facial features had deteriorated, and the characters' skin looked like wax - this is truly bad news!
I've noticed an increase in waxy skin too. I think it is due to worse compatibility with realistic LORAs in pure 2511. The mole issue I think is related to what your sampler and step configuration is, as I believe that is avoidable.
Give v21 a try.
thanks for sharing... good work..
but the consistency is a big problems. I tried many version, v19,v20 has a serious problem.
i use the standart qwen 2509 workflow. i am not sure it is the right one..
v21 appears unable to do male genitals (unless you want them with labia).
FIrst of all, big fan here. V18, from my tests are the best in terms of consistency and also realism. But don't understand NSFW concepts as well as 20/21. I just wish you didn't add faceswap in new models. It's like giving mixed signals to the checkpoint. Like, I want you to keep consistency, but also want you to change face...
thanks for sharing... good work..
but the consistency is a big problems. I tried many version, v19,v20 has a serious problem.
i use the standart qwen 2509 workflow. i am not sure it is the right one..
Please share your workflow showing this problem.
v21 appears unable to do male genitals (unless you want them with labia).
Please share your workflow showing this problem.
FIrst of all, big fan here. V18, from my tests are the best in terms of consistency and also realism. But don't understand NSFW concepts as well as 20/21. I just wish you didn't add faceswap in new models. It's like giving mixed signals to the checkpoint. Like, I want you to keep consistency, but also want you to change face...
Please share your workflow showing this problem.
After reading the comments above, I finally know how to describe my feelings about this V20 version. At first, I thought it was because the character skin colors were more vibrant and the skin was more defined. And indeed, the texture of the character skins can now be described as waxy.
FIrst of all, big fan here. V18, from my tests are the best in terms of consistency and also realism. But don't understand NSFW concepts as well as 20/21. I just wish you didn't add faceswap in new models. It's like giving mixed signals to the checkpoint. Like, I want you to keep consistency, but also want you to change face...
Agreed，
FIrst of all, big fan here. V18, from my tests are the best in terms of consistency and also realism. But don't understand NSFW concepts as well as 20/21. I just wish you didn't add faceswap in new models. It's like giving mixed signals to the checkpoint. Like, I want you to keep consistency, but also want you to change face...
Agreed，
I'm still awaiting an example where this is happening, then I can test with or without the (very small amount of) BestFaceSwap LORA to see if it makes a difference.
Just adding in a small touch of the FaceSwap LORA isn't embedding a prompt to "change the face". It just tweaks some of the weights to better support facial consistency when prompted to face swap. I only add in a very small amount, designed to not overfit the merge to only face swapping use cases. Until it can be demonstrated with a back to back example, claiming it is 0.2 of the BestFaceSwap LORA is just conjecture.
FIrst of all, big fan here. V18, from my tests are the best in terms of consistency and also realism. But don't understand NSFW concepts as well as 20/21. I just wish you didn't add faceswap in new models. It's like giving mixed signals to the checkpoint. Like, I want you to keep consistency, but also want you to change face...
Agreed，
I'm still awaiting an example where this is happening, then I can test with or without the (very small amount of) BestFaceSwap LORA to see if it makes a difference.
Just adding in a small touch of the FaceSwap LORA isn't embedding a prompt to "change the face". It just tweaks some of the weights to better support facial consistency when prompted to face swap. I only add in a very small amount, designed to not overfit the merge to only face swapping use cases. Until it can be demonstrated with a back to back example, claiming it is 0.2 of the BestFaceSwap LORA is just conjecture.
I posted some examples here how v21 smoothed out all the details and changes the face consistency. If you look to the side where the roots are, all the details are gone. https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/discussions/244#696e495e500d013f644022a8
I've looked it over and came to some conclusions in that thread.
I've noticed an increase in waxy skin too. I think it is due to worse compatibility with realistic LORAs in pure 2511. The mole issue I think is related to what your sampler and step configuration is, as I believe that is avoidable.
I tried 4, 6, and 8 steps, but imho, increasing them adds "waxiness". It's been a while since I visited here, didn't notice there was a whole discussion going on. Actually, I just kept using version 19 and am currently downloading version 22. Thank you for the feedback.