Bug: Does not work on models generated by dreambooth
AndrewKrill opened this issue · comments
Regardless of prompt, input image, the result is always the same. Usually a near-transparent mask. Will be all light grey/white if using the "grey" output mode.
Could you provide your settings?
That will include the image, prompt, indices and the cross-attention layer.
I initially thought this was an issue with VXA script. I dug into the code to debug and compared with DAAM script.
I think this issue is actually with certain trained models. I believe dreambooth may be to blame, but I need to fix my dreambooth install to debug. I would like to dig in and figure out why this is happening.
I will close this bug out for VXA. Posting the following to share my findings:
If I use the standard SD model I get results I would expect.
Index 1 highlights clouds. If I add a comma, then index 2 highlights clouds.
If I use a model from dreambooth, the first token is garbage regardless.
I only get the expected result if I add a garbage token then use the second index.
I tried out DAAM script with img2img, same behavior.
Expected results with SD model:
First token is garbage for a trained model: