Find AI Tools
Stable Attribution screenshot

What is Stable Attribution?

Stable Attribution was a free tool built at Chroma that decoded an image generated by an AI model into the most similar examples from the data the model was trained on. The goal was to restore credit to the artists and photographers whose work was scraped to train image generators such as Stable Diffusion, by surfacing the original source images and creators. The project has since been sunset by its authors, and the interactive lookup is no longer operational, though the explanatory FAQ remains online. It was created by Jeff Huber and Anton Troynikov at Chroma.

Key Features

Training data lookup

Decodes an AI-generated image into the most similar examples from the dataset the model was trained with.

Similarity search engine

Uses knowledge of the Stable Diffusion model weights to augment a similarity search across the training images.

Source image surfacing

Returns the training images judged most likely to have influenced a given generated image.

Artist crediting goal

Aimed to assign attribution back to the original artist or creator of each source image.

No data claims

The authors stated they did not claim rights to any uploaded or generated images and would not train models on them.

Community identification

Invited the public to help identify artists in the discovered source images.

Open research direction

Documented limitations of Version 1 and ongoing research into broader generative model attribution.

Pros & Cons

Advantages

  • The underlying idea addressed a genuine concern around consent, credit and compensation for artists whose work trained AI image models.
  • The tool was free to use, with no account, subscription or payment required.
  • The authors were transparent about how the similarity-based attribution worked and about its limitations.
  • They publicly committed not to claim rights to uploaded images or to train any models on them.
  • The FAQ openly acknowledged that Version 1 was imperfect rather than overstating accuracy.

Limitations

  • The project has been sunset by its authors and the interactive image lookup is no longer operational, so the homepage now shows only a farewell message.
  • Attribution was based on visual similarity rather than confirmed provenance, so matches were estimates and could be wrong.
  • Version 1 was acknowledged to be imperfect due to noisy training processes and dataset errors.
  • There is no ongoing support, roadmap or active maintenance, as the team moved on to other work at Chroma.

Use Cases

Artists checking whether their work appeared among the closest training images for a given AI generation.

Researchers exploring how training data influences the output of diffusion-based image models.

Writers and journalists illustrating the data-provenance and copyright debate around generative AI.

Designers and creators investigating the likely sources behind a specific AI-generated image.

Educators demonstrating how AI image attribution and similarity search work in practice.