In-painting improvements

• In-painting now leverages the Qwen VL vision model to more accurately generate the in-painting mask. This should result in more accurate results and a better overall experience.

API

• Revised the response format of the /api/v1/models/compatibility_mapping and /api/v1/models/traits endpoints to better represent their data structure. These endpoints have been added to the docs and Postman has been updated.

App

• Fixed an issue where users using multi-modal vision models may have seen failed inference after multiple requests.