EditAgent expands tools and verification for agentic image editing
Original title: EditAgent: Rethinking Agentic Image Editing
Experts preferred parallel listwise judging over sequential refinement in 78.5% of decided comparisons.
EditAgent routes each editing step to general editors and specialized methods and attaches a verification record to every candidate. With those records, off-the-shelf vision-language judges outperform finetuned reward models. A general-purpose editor produced no usable result in over 40% of expert-annotated cases. EditAgent improved two open-weight editors on three benchmarks and can generate training data.
- Candidates are generated in parallel and judged in one listwise call.
- Listwise judging required no more than half the judge calls.
- No tested judge was reliable across every edit type.
- EditAgent also serves as a data engine for training stronger editors.
From the original article
1 min read at huggingface.co
Agentic image editors decompose complex requests into steps, but when a step fails, they mostly retry with the same narrow set of tools and a judge that scores each output by eye directly. We argue that such agents are bounded by two spaces: the action space of what the agent can do, and the observation space of what its judge can see.
By Haiyang Xu, Tianyu Wang, Jialu Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.