Why it’s interesting
Surya Narreddi used reinforcement learning to train Qwen to generate editable painting code, presenting the aesthetic rewards, failed attempts, and watercolor results as a viewable research artwork.
Instead of producing an image from one prompt, the system turns aesthetic preference into a reward and learns through repeated code revisions.
How to try it
Open the project page and scroll from the training diagram through the reward design and failures to the final watercolor paintings.
Use the official page or demonstration as the current reference. Do not assume that a staged installation, limited experiment, account-gated demo, subscription feature, or physical product is available everywhere.
Requirements and sources
This entry credits Surya Narreddi and keeps “Training AI to Paint with Code” as the original title. It is free to view or try and available as an immediate browser or media experience. Pricing, access, regional availability, accounts, hardware, and service support can change, so verify them on the [first-party source](https://surya.website/rling-qwen-to-paint-with-code) before trying or buying. The description stays within the documented AI role and does not treat promotional claims as independent testing.