Visual interpretability agent

Akin to MAIA equip a VLM, e.g., Qwen 2.5 VLM or Gemma 3 VLM or some model behind an API with tools for (basically what we provide in our demos), e.g.,

• generating images using SDXL/FLUX,
• visualizing features active during the generation,
• for a given feature, compute images for which this feature was on,
• performing feature interventions by turning on / off features

and prompt it / provide it with additional scaffolding to come up with hypotheses about features as well as functionality to validate those hypotheses. Our current goal for this agent is not so much only coming up with explanations of features but more to perform targeted interventions on generated images as is required for our recent representation-based editing benchmark.

We provide extensive resources including descriptions and code for our SDXL/FLUX SAE features here.