TILT Framework for Compositional Text-to-Image Diffusion
May 15, 2026
TILT improves prompt faithfulness in diffusion models using a training-free, model-intrinsic reward. It utilizes a KL-constrained objective with a closed-form tilted target distribution to resolve concept overlap failures during test-time sampling.
HOW THIS AFFECTS YOU
●
researcherYou can achieve better compositional adherence without additional supervised training.
●
designerThis allows for more precise control over complex multi-subject image generation.