• You can provide between 1 and 5 inputs, these can either be an uploaded image a text prompt or a url to an image file.
  • The order of the inputs shouldn’t matter, any images will be centre cropped before use.
  • Each input has an individual strength parameter which controls how big an influence it has on the output.
  • The model was not trained using text and can not interpret complex text prompts.
  • Using only text prompts doesn’t work well, make sure there is at least one image or URL to an image.
  • The parameters on the bottom row such as cfg scale do the same as for a normal Stable Diffusion model.
  • Balancing the different inputs requires tweaking of the strengths, I suggest getting the right balance for a small number of samples and with few steps until you’re happy with the result then increase the steps for better quality.
  • Outputs are 640×640 by default.
  • If you want to run locally see the instruction on the Model Card.

How does this work?

This model is based on the Stable Diffusion Image Variations model but it has been fined tuned to take multiple CLIP image embeddings. During training, up to 5 random crops were taken from the training images and the CLIP image embeddings were computed, these were then concatenated and used as the conditioning for the model. At inference time we can combine the image embeddings from multiple images to mix their concepts (and we can also use the text encoder to add text concepts too).

The model was trained on a subset of LAION Improved Aesthetics at a resolution of 640×640 and was trained using 8xA100 GPUs on Lambda GPU Cloud.

Leave a Reply

Help us find great AI content


Never miss a thing! Sign up for our AI Hackr newsletter to stay updated.


AI curated tools and resources. Find the best AI tools, reports, research entries, writing assistants, chrome extensions and GPT tools.