2025/09/02 by Yiflach, Sapir Esther, Atzmon, Yuval, Chechik, Gal
#Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences
paper · doi:10.48550/arxiv.2509.02295
Text-to-image diffusion models can generate stunning visuals, yet they often fail at tasks children find trivial--like placing a dog to the right of a teddy bear rather than to the left. When combinations get more unusual--a giraffe above an airplane--these failures become even more pronounced. Existing methods attempt to fix these spatial reasoning failures through model fine-tuning or test-time optimization with handcrafted losses that are suboptimal. Rather than imposing our assumptions about spatial encoding, we propose learning these objectives directly from the model's internal representations. We introduce Learn-to-Steer, a novel framework that learns data-driven objectives for test-time optimization rather than handcrafting them. Our key insight is to train a lightweight classifier that decodes spatial relationships from the diffusion model's cross-attention maps, then deploy this classifier as a learned loss function during inference. Training such classifiers poses a surprising challenge: they can take shortcuts by detecting linguistic traces rather than learning true spatial patterns. We solve this with a dual-inversion strategy that enforces geometric understanding. Our method dramatically improves spatial accuracy: from 0.20 to 0.61 on FLUX.1-dev and from 0.07 to 0.54 on SD2.1 across standard benchmarks. Moreover, our approach generalizes to multiple relations and significantly improves accuracy.