is there inpainiting dataset and parameters example provided for xl training?

[fire2323]

What API design would you like to have changed or added to the library? Why?

What use case would this enable or better enable? Can you give us a code example?

Hi patil-suraj @patil-suraj , appreciated for the convenient script ! Is there any code example and dataset example to run the script: https://github.com/huggingface/diffusers/blob/inpainting-script/examples/inpainting/train_inpainting_sdxl.py ?

[fire2323]

@sayakpaul could you also check this issue?

[sayakpaul]

This is not an issue of the diffusers library and should go into discussions.

[sayakpaul]

To answer your question, unfortunately no.

[fire2323]

could provide anyone by the script author? There seems to have problems to run this script with my cutomized dataset. The script gets stuck when loading the batch data.

[fire2323]

Got. Thanks for the information.

[fire2323]

BTW, which techniques do you think have better results ?

[asomoza]

it depends on the model but for the ones that support it that would be differential diffusion which is a training free solution, also I use a lot the controlnet, I did some guides and spaces with those and people really like them, specially the outpainting one that was made by a community user.

https://huggingface.co/spaces/fffiloni/diffusers-image-outpaint
https://huggingface.co/spaces/OzzyGT/diffusers-fast-inpaint
https://huggingface.co/spaces/OzzyGT/diffusers-image-fill

for SD 3 that would be also the controlnet but since the model isn't used that much, is not that well know but it's really good, you can also use differential diffusion with SD 3 and SD 3.5.

for Flux there's the Flux Fill solution which is a whole model finetune but people seems to like it a lot if you have the resources to run it.

As you can see, none of the newer models use or trained an inpainting model, unless you consider the Flux one that but Flux it is in a league on their own, in the sense of quality, open source and the model itself, for example the depth and canny are also a whole model finetune instead of a smaller controlnet.

[fire2323]

OK, THX !

[fire2323]

i would like to have a try