-
|
I have Checkpoints that generate artifacts and they recommend using VAEs to avoid it, but I don't know how to choose a specific one. In the VAE folder I have 2, and although they are in the corresponding folder, I still see artifacts. In total I have come across 3 checkpoints with this artifact problem. The checkpoint of the preview I uploaded is this: https://civitai.com/models/120090?modelVersionId=268240 So I was wondering how to solve this, since I'm finding more and more checkpoints that need VAEs to function properly. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 11 replies
-
|
Is another Sampler not suitable for you or gives the same result? |
Beta Was this translation helpful? Give feedback.
-
|
I haven't tried this model, so I'm not sure if this would help - but as poor7 pointed out, perhaps the sampler will make a difference. I see that in most of the images that the model's creator posted it looks like they're using "Euler a" as the sampler, while the default sampler for Fooocus is "dpmpp_2m_sde_gpu". You might try using the settings the model's creator is using and see if that helps. They're using Euler a which is called "euler_ancestral" in Fooocus, as well as a sampling step count of 20. To change the Sampler in Fooocus you can do that under the "Advanced" tab >> "Developer Debug Mode" >> "Sampler". You can change it to "euler_ancestral" to see if that helps. You can also change the step count under Developer Debug Mode and "Forced Overwrite of Sampling Step". Hopefully that will help with the issue. :) |
Beta Was this translation helpful? Give feedback.




@Madxis No, I unchecked all the styles as I didn't want it to effect the outcome of the image. There are definitely still artifacts but it's not as bad. I do think there's a way to change the VAE but hopefully there are also internal settings that can be found to fix the issue. It seams that the Model is perhaps a little "over trained"? I'm not sure, but I've never seen that issue within Fooocus before, so I'm sure it must have something to do with the model or how it was trained. Hopefully you'll find a better solution or a model that works a little bit easier with Fooocus. :)