SALL-E
This is the repo containing SALL-E a fine-tuned Stable Diffusion V1.5 model on DALLE-3 generated data. DALL-E was named after the movie WALL-E and the artist Salvador Dali, SALL-E on the other hand is a mix of Stable Diffusion, WALL-E, and Salvador!
- The model is released using
.safetensorsand was tested on ComfyUI interface and Automatic1111. You could also use it with Diffusers. We recommend using DPM++ 3M SDE as a sampler and Karras as a scheduler, guidance scale of 6 and around 20 steps of denoising. When using the HiRes Fix we do 15 steps in the first denoising stage and another 15 after upscaling the latent using nearest-exact option. - Our testing reveals significant improvement in the generated samples in terms of textual alignment of the generations and aesthetics. We tested the model on prompts generated using ChatGPT4, CivitAI and PromptHero.
- Future plans include releasing a LORA SDV1.5 and SDXL Fine-tune and LORA as well.
- Training setup and detailed will be released soon.
Sample Images
Contributions:
- If you want to help in this project, please reach out with any possible advice on training setups and tricks you have learned during your model fine-tuning! This would be very helpful for training more powerful models in the future. Just open an issue and let's have a conversation!