Adobe's EditAgent routes image edits across editors with evidence-based listwise judging
adobe · hf · 2026-10-30
Adobe introduces EditAgent, rethinking agentic image editing for reliability and efficiency.
- Action space: instead of retrying the same tools blindly, it plans, locates the edit region, and routes steps to diverse editors (general, local, tone, style), producing multiple candidates in parallel.
- Observation space: rather than eyeballing outputs or finetuned reward models, an off-the-shelf VLM performs listwise judging in one call using evidence—change maps, before/after close-ups, target masks, luminance/color stats, geometry, and code-verified facts.
- Results: expert-preferred edits with fewer judge calls; evidence lifts every VLM judge, beating finetuned reward models, and the system doubles as a data engine for training stronger editing models.
More from Multimodal
- Midjourney v8.2 Naruto-style renders shared with reusable sref code 2565685973 — azed_ai · 2026-10-30
- Video Model Shootout: Pika H3 Flex vs Three Distilled H3 Recipes on Character Continuity — rileygstaliger · 2026-10-30
- seriesmith: MIT-Licensed Tool Generates 22-Minute AI Episodes From Scripts With Character Consistency — bgstratt · 2026-10-30
- Anime-Style Short Film Made With Seedance and Fish Audio Goes Viral — YumetsuzuriAnime · 2026-10-30
- Short film with an all-AI cast is indistinguishable from real actors, says viewer — TotGaH · 2026-10-30
- ElevenLabs SFX v3 handles multi-sound prompts that broke earlier AI SFX tools — HeyAmit_ · 2026-10-30