Persistent Plastic Skin Artifacts in Qwen Image Edit
It’s good, but the plastic-looking issue is still very severe.
Of course, this can be seen as a chronic problem of Qwen Image Edit, but could this be a sampling-related issue?
There are too many cases where the face changes drastically or the skin is rendered with an overly plastic appearance.
I'm uploading a v23 now. However, I think there is a bit of varying expectations, and perhaps a bit of exaggeration when I hear claims of "face changes drastically", "chronic", "very severe" and "overly plastic appearance" (see this thread: https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/discussions/244). Prompts matter a bunch too, so don't use terms like "don't change the face" and if you want more realism, try terms like "photograph" (and if you want noise, try terms like "noisy 2000s smartphone candid digital photo" or something similar).
Also, getting a bit tired of these big claims without any examples or workflows to back them up. Honestly, I'm going to stop responding to posts without examples or workflows, because there isn't anything concrete I can work from.
Btw, congratz on v23. I'm testing and it's amazing! No problems at all for now. You hit the right spot!
Hearing that, I immediately started downloading version 23.
Just like you said, the prompt and euler_ancestral were effective, and I’m already excited to see how much has been improved in version 23.
Hey, I just tried it! Seriously, v23 with the Euler_ancestral sampler delivers the best performance I’ve seen so far!
Version 23 is the best yet. If you dont want plastic look, use bong tangent sampler, like someone here mention. works great.
用z-image在轻微重绘一次就好了。
I'm uploading a v23 now. However, I think there is a bit of varying expectations, and perhaps a bit of exaggeration when I hear claims of "face changes drastically", "chronic", "very severe" and "overly plastic appearance" (see this thread: https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/discussions/244). Prompts matter a bunch too, so don't use terms like "don't change the face" and if you want more realism, try terms like "photograph" (and if you want noise, try terms like "noisy 2000s smartphone candid digital photo" or something similar).
Also, getting a bit tired of these big claims without any examples or workflows to back them up. Honestly, I'm going to stop responding to posts without examples or workflows, because there isn't anything concrete I can work from.
I’ve never heard something so spot-on. I totally get what you’re saying, it’s exhausting. I felt like I had to respond to this. You’re absolutely right. It makes no sense not to send the workflow.