Scale of KD-feature loss for SD inpainting 1.5

[Bikesuffer]

Hi there,

I am trying to distill the Unet in SD inpainting 1.5 to a smaller Unet by using your code. (I replaced the pipeline to inpainting and the input data)
I have trained for 130K steps with batch size 64.
Right now the kd_feat_loss is around 20.

I am wondering what kd_feat_loss you have when you finished distill the Unet in your experiment?

Thank you.

[bokyeong1015]

Hi, thanks for utilizing our work, glad to know that 😊
Although we haven't attempted inpainting experiments, we hope the following information can be helpful.

---

Here is a loss curve from our code for text-to-image synthesis, with SD-v1.4 and batch size 64 (= gradient accumulation 4 x mini batch size 16), plotted with 500-point moving average:

• The scale of KD feature loss ≫ The scale of KD output loss and SD task loss
  • As we described in our paper, we didn’t try hyperparameter tuning for loss weights, but it empirically worked well in our experiments.
• Losses are not directly correlated with the final generation scores (FID/IS/CLIP score), especially in later iterations. In other words, lower losses did not necessarily result in better generation scores.
• If you want to verify the learning process, we suggest examining the final metrics and/or visual examples. Nevertheless, the losses should decrease during initial iterations.

[bokyeong1015]

Please understand that we've changed the name of this issue, 'Batch Size' -> 'Scale of KD-feature loss for SD inpainting 1.5', to clarify the topic and make it easier for people to find in the future.

[Bikesuffer]

Thanks a lot for the information.

[yajieC]

hello, does this method work for SD inpainting 1.5?

[bokyeong1015]

Hi, @yajieC
We haven't tried it, but we believe our models can be used after finetuning for SD-inpainting.

Our models are compressed from SD-v1.4, and SD-v1.x models share the same architecture (with different training recipes); SD-inpainting was based on SD-v1 backbone.

[Bikesuffer]

hello, does this method work for SD inpainting 1.5?

Yes, it worked for me.
I have successfully distill the unet in sd inpainting 1.5 to a smaller Unet
I would say the SD_base model distilled with batch size 256(I call it IP_Base_256) generate best result for me.

[bokyeong1015]

Thanks for sharing the above and this good news! Happy to know you are okay with the inpainting results using our approach :) Could we ask if you have plans to release your models and/or code?

---

Edit: sorry for initial misunderstanding, you've clarified that "distill the unet in sd inpainting 1.5 to a smaller Unet", which means (Teacher, Student) = (SD-inpainting 1.5, BK-SDM modified using additional input channels) <- please let us know if this is incorrect updated. Thanks again for sharing! @Bikesuffer

[Bikesuffer]

Thanks for sharing the above and this good news! Happy to know you are okay with the inpainting results using our approach :) Could we ask if you have plans to release your models and/or code?

---

Edit: sorry for initial misunderstanding, you've clarified that "distill the unet in sd inpainting 1.5 to a smaller Unet", which means (Teacher, Student) = (SD-inpainting 1.5, BK-SDM) <- please let us know if this is incorrect. Thanks again for sharing! @Bikesuffer

Hi actually the student is a modified version of bk sdm since the input of unet in inpainting pipeline is 9 channel. But all the anchor points for calculating the loss are the same as bk sdm.

[bokyeong1015]

Thanks for the clarification, and we've updated the student description in the above :)

[yajieC]

hi, I tried this method, but found that the performance was very poor. My experimental configuration was to train on laion_11k data for 10k steps, and the unet is bk_tiny. And I also replaced the pipeline to inpainting and the input data. I would like to ask you for any good suggestions, thanks.

[bokyeong1015]

@yajieC Thanks for your inquiry. We would like to address this in a separate discussion for making it easier for future readers to find, because it seems a different topic. Please kindly refer to our response at that link.