About

A Cog implementation of ostris/ai-toolkit for training StableDiffusion3.5-Large LoRAs.

Follow me on Twitter/X @lucataco93

How to use

In the TRAIN tab (between README and VERSIONS) you’ll find the parameters you need to train a LoRA

For destination - select/create an empty Replicate model location to store your LoRAs. (Ex: lucataco/flux-loras)

For images - upload a zip or tar file of images for training. Images should be in jpg, jpeg, or png format. You may upload the images alone without captions or include a corresponding txt file for each image. If using captions, name each txt file to match its image (e.g., for my-photo.jpg, the caption file should be named my-photo.txt and contain the caption text).

For hf_token - use your Hugging face token to access the Flux-Dev weights for training. Make sure the Access Token has write permissions to upload models to your Hugging face account.

For steps - select a value from 700-4000

By default the trigger word is: ‘TOK’. The other steps are optional

Example training run

Below is a training run of a dog Queso (Trigger word QSO) using 12 images and separate caption txt files at 1000 steps that took 20min:

How to run your LoRA

Once your LoRA is in Huggingface, go to the model and grab the full URL to the LoRA and use it with the model: lucataco/stable-diffusion-3.5-large-lora

For example, using the example training run Queso the following is an inference run of that LoRA:

Licensing and commercial use

If you generate images on Replicate with StableDiffusion3.5-Large models and their fine-tunes, then you can use the images commercially.

If you download the weights off Replicate and generate images on your own computer, you can’t use the images commercially.