Flux1: using sd-scripts trained LORA with diffusers

[ddpasa]

I have a really nice Lora that I trained with sd-scripts. It works very well with SD-Forge. Unfortunately SD-Forge does not seem like it is being developed actively, and I want to try a simpler alternative. Is it possible to load the same Lora with the Huggingface diffusers library? Would it need some kind of conversion to work properly.

[ddpasa]

Ok, after doing some research I found the following:

https://github.com/kohya-ss/sd-scripts/blob/sd3/networks/convert_flux_lora.py is used for converting from sd-scripts format to the diffusers format. however, it does not support either of the text encoders (neither clip nor t5xx) and just converts the transformer.

@kohya-ss is that accurate?

[IAmMorrow]

why don't you use comfyui ? works well out of the box and is pretty much the standard currently. If you don't like linking nodes, swarm ui is another "more traditional" frontend for comfyui

[ddpasa]

why don't you use comfyui ? works well out of the box and is pretty much the standard currently. If you don't like linking nodes, swarm ui is another "more traditional" frontend for comfyui

I like having a simple python script that is easy to automate. ComfyUI is too much. I'll check Swarm, but I'm looking for a simple lightweight thing.

[kohya-ss]

Unfortunately, convert_flux_lora.py does not support converting LoRA modules for Text Encoders.