Nerogar
[Feature] Able to set Conv layer padding mode
Hello folks,
Although there are some ways to generate tileable textures, like setting Conv layer padding mode to "circle" when generating, however, this kind of operation would greatly impact results with a bad influence. Hence, during training, if we are able to set the padding mode to "circle" by natural, then the trained model would have a good performance at generating tiling images when enable "tiling" feature in A1111.
This feature is often lacking in current finetuning tools on the market, which would make OneTrainer unique if it provided this feature during training.
Best regards.
Reference: https://github.com/huggingface/diffusers/issues/2633#issuecomment-1676629872
Yes. This is an easy to add feature. I will add it. But right now all my time is spent adding stable cascade support.
Yes. This is an easy to add feature. I will add it. But right now all my time is spent adding stable cascade support.
Thanks! Can't wait to see it. Btw, what is stable cascade? May I have some link to get familiar with it? Thanks
new t2i model from sai https://stability.ai/news/introducing-stable-cascade, and there should be another one on the 22nd
new t2i model from sai https://stability.ai/news/introducing-stable-cascade, and there should be another one on the 22nd
Thanks!
When replacing the padding mode for SDXL, there are 2 TE, we need to replace all of the components of SDXL like: targets = [pipeline.vae, pipeline.text_encoder, pipeline.text_encoder_2, pipeline.unet]
Completely forgot about this. I added the setting now to the train tab. I don't have a good dataset to test it myself, so let me know how it goes.
When replacing the padding mode for SDXL, there are 2 TE, we need to replace all of the components of SDXL like: targets = [pipeline.vae, pipeline.text_encoder, pipeline.text_encoder_2, pipeline.unet]
The text encoders don't use conv layers, so this doesn't apply.
Hi guys @mx @Nerogar
Just check the codes, seems we only applied the circular padding mode but without changing the behavior of conv2d. Could you please check this repo as reference please? Since it helps to generate seamless image. Thank you very much.
Ref: https://github.com/spinagon/ComfyUI-seamless-tiling/blob/master/SeamlessTile.py#L46