Training code for SDXL

[youngwanLEE]

Hi,

I'm impressed by your amazing work !!

Do you plan to open the training code for SDXL?

It would be helpful for the open-source community.

[G-U-N]

Hey @youngwanLEE. Definitely will release the training code for SDXL! So many things occupy my time currently, I will upload them later. Thank your for the attention!

[lovejing0306]

+1

[xiao2mo]

waiting for it.

[G-U-N]

Hi, guys. I have been working on Stable Diffusion 3 in the past few days. I have uploaded all the training scripts for it. I hope they will be helpful. The training scripts for SDXL will also be uploaded after verifying everything is okay.

[xiao2mo]

Hi, guys. I have been working on Stable Diffusion 3 in the past few days. I have uploaded all the training scripts for it. I hope they will be helpful. The training scripts for SDXL will also be uploaded after verifying everything is okay.

SDXL waiting for it ^_^ , any time plan?

[G-U-N]

I hope it will be quick. I am trying to apply for some GPUs for verifying before uploading.

[bengen-y]

looking forward to it ！！！

[xiao2mo]

looking forward to it ！！！

[xiao2mo]

looking forward to it ！！！

[xiao2mo]

looking forward to it ！！！

[G-U-N]

@xiao2mo . Thanks for the interest! Will upload it asap.

[xiao2mo]

asap

😁

[xiao2mo]

@xiao2mo . Thanks for the interest! Will upload it asap. hi。

[xiao2mo]

looking forward to it ！！！

[youngwanLEE]

@G-U-N While waiting for your code, I have tried implementing PCM-SDXL. I wonder if the discriminator for SDXL has the same architecture as SD-v1.5.

[George0726]

I have trained PCM on my fine-tuned 9-channel outpainting SDXL. I found that 8-step is much better than 4-step that is different from Fig 7. in your paper. I am not sure if it is owing to the outpainting tasks or the convergence of my model. I wonder if you can reveal your SDXL based architecture so that I can verify my current experiments.

[G-U-N]

Hey all, SORRY for the late update. I just got access to my original code on the GPU server.

The LoRA training script is now available at run.sh.

[G-U-N]

Typically 2k iterations is enough to see clear improvements. Yet I have no resources to retrain everything. LMK if you encounter any issues.

Best, Fu-Yun

[G-U-N]

500 iter, 1000 iter, 1500 iter, 2000 iter with a batch size 40. Longer training makes it better.

[youngwanLEE]

I successfully reproduced PCM-SDXL and implemented PCM-KOALA models using KOALA U-Net backbones.

I really appreciate your open-source effort.