So how to use this format with stable diffusion pipline? especial the .safetensors of lora about safetensors HOT 9 CLOSED

it seems like this will soon be merged into diffusers library huggingface/diffusers#2448 and in the meantime this script may be useful: https://github.com/haofanwang/Lora-for-Diffusers/blob/main/format_convert.py

from safetensors.

Well, LoRA are just regular weights, so there's nothing blocking from using safetensors for those.

No offense but both your links are code I have never heard about. I'm happy to help provide guidance in creating safetensors lora support in all of them, but I cannot reasonably create all the PRs in all the repos. Just understanding what's going on is going to take me too long.

But basically search of torch.load/torch.save and replace with safetensors.torch.save_file/load_file and you should be good.

from safetensors.

Do you mean this ?

https://huggingface.co/docs/diffusers/training/lora#inference

from safetensors.

I created a PR to enable support in diffusers. Is that what you meant ?

huggingface/diffusers#2448

from safetensors.

I created a PR to enable support in diffusers. Is that what you meant ?

huggingface/diffusers#2448

Absoultely not! Training lora has at least 4 methods, https://github.com/kohya-ss/sd-scripts; https://github.com/Akegarasu/lora-scripts; webui extention; https://github.com/huggingface/diffusers/tree/main/examples/dreambooth#inference-1. Thought them are in format of safetensors, but they are not same inside!

from safetensors.

Well, LoRA are just regular weights, so there's nothing blocking from using safetensors for those.

No offense but both your links are code I have never heard about. I'm happy to help provide guidance in creating safetensors lora support in all of them, but I cannot reasonably create all the PRs in all the repos. Just understanding what's going on is going to take me too long.

But basically search of torch.load/torch.save and replace with safetensors.torch.save_file/load_file and you should be good.

I wish it was. In https://civitai.com/, most of safetensors of lora cant't be loaded in diffusers.

from safetensors.

You mean these kind of checkpoints ? https://civitai.com/models/5373/makima-chainsaw-man-lora

It seems like this is a Lora and should be already usable in SD, no ?

from safetensors.

@Narsil's PR adds safetensor's to our current saving and loading of lora weights. Happy to discuss adding more lora/safetensors functionality but we would need more details and would recommend a separate, more detailed issue on diffusers :)

from safetensors.

Closing the issue. Feel free to reopen

from safetensors.