Grounded image inpaint - not in the demo

[stared]

On the GLIGEN website, there is an example of inpainting that inserts images rather than generating them from a text prompt.

However, I am unable to find this option in the demo. Is it available somehow?

[haotian-liu]

It not available in demo yet (due to it potentially can have multiple objects), while you can try it with our inference script: https://github.com/gligen/GLIGEN/blob/master/gligen_inference.py#L367

[stared]

@haotian-liu Thank you - I figured it out, and it works - see #22 . Would you be interested in adding it?

I had trouble relating it to the demo code (in gradio), though. The code seems to be different from the one in https://github.com/gligen/GLIGEN/blob/master/gligen_inference.py. Does it work the same way (just duplication written in a different way) or (as it seems) are there some other improvements, e.g. related to xformers?

[PrajwalCogniac]

Hey @haotian-liu @stared Even I want to try the same application ! How did you manage to do it ? As there is no checkpoint file available ? All there is a .bin file from HUB how did you use that ? Any help to run the inference in this regard will be helpful

[avivbur]

@PrajwalCogniac You can refer to the Download GLIGEN Models table in the README.md file - I believe the model you need is Inpainting | Box+Text+Image

[CREED404]

@avivbur does I still need to download the checkpoint file if I am using the Gligen demo app? doesn't it download all files automatically?