SearchGen: Search-Augmented Visual Generation
TIGER-Lab · hf · 2026-07-15
This work introduces SearchGen-20K and SearchGen-Bench to explore search-augmented visual generation.
Key findings:
- Visual generation models confidently hallucinate novel entities and long-tail events outside their training data.
- On SearchGen-Bench, frontier open-source generators score only 21–28/100, revealing a larger capability gap than existing benchmarks.
- Directly adding search tools isn't always effective; naive retrieval can introduce noise into prompts.
- The authors argue the key is a model's "internalizable knowledge boundary" and propose a teach-then-search co-training framework to progressively discover and leverage it.
The paper also releases reproducible tool-augmented benchmarks, corpora, and retrieval datasets.
Related event: SearchGen: Enhancing Visual Generation via LLM Search(2 posts)→
More from Multimodal
- Storyboard-first workflows are making AI dance videos and influencers more consistent — aftahi_ai · 2026-07-22
- Interactive video should be judged by responsiveness, not just frame quality — Soggy_Limit8864 · 2026-07-22
- Runpod MCP and Claude help spin up image and video generation workflows — 802high · 2026-07-22
- Midjourney prompt turns a bee into a glitching pixel explosion — michaelrabone · 2026-07-22
- A physics reward can improve video generation without creating a real physics engine — Dapper-Drawer4546 · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22