The authors built SpatialCritic, an 8-billion-parameter reward model that verifies whether generated images satisfy spatial prompts.
Spatial rewards and staged training improve image layout accuracy
The authors combine a tool-using image critic with a capability-based curriculum to improve spatial prompt adherence across two image generators.
Chinese Tech
Xiaofeng Tan · Jiaxin Lin · Haobo Yuan · Tiancheng Shen · Zihan Su · Rui Chen · +2 more
Southeast University · Tencent HY · UC Merced · Tsinghua University
Research Digest··2 min read
Tan et al.
Why this paper
From Tencent HY and 3 others
In one line
An agentic reward model and curriculum learning together improve spatial accuracy in text-to-image generation.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 8B)
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§