Skip to content

gnguralnick/stable-diffusion-fine-tuning

Repository files navigation

stable-diffusion-fine-tuning

This project was completed as part of CSC413 at the University of Toronto.

The training-scripts directory contains the code to train the fine-tuned models, with a subdirectory for each fine-tuning method.

The fine-tuned models directory contains some of the final outputs of the fine-tuning methods, at least for textual inversion and LoRA. The outputs for Dreambooth were not included due to their size.

The target-images directory contains the original target images that the fine-tuned models were trained with.

The generated-images directory contains the images generated by each fine-tuned model. They are grouped by fine-tuning method and by prompt type, where "basic" is the default prompt "a photo of <*>" while the other groups, like "forest", are the more complex prompts like "a photo of <*> in a forest".

The target-complex-images directory contains images generated by each fine-tuned model for the more complex prompts that did not include the placeholder token (e.g., "a photo of a in a forest"). These images were used to evaluate how well the fine-tuned models were able to incorporate other features of prompts aside from just the learned concepts.

The evaluation-scripts directory contains several scripts for running our evaluation process, including run_inference for generating the actual output images, clip_distance for performing the CLIP similarity comparisons, and fid for performing the Fréchet inception distance calculation.

Finally, evaluation-results contains the actual results from our evaluations.

All of these scripts can be run provided that the command pip install -r requirements.txt has been run. Python 3.10 is required due to the use of PyTorch 2.0.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages