TILT boosts diffusion model compositionality with intrinsic reward

Compositional text-to-image generation remains challenging for diffusion models, which often fail to faithfully combine multiple concepts from a prompt. A new paper from Debottam Dutta and colleagues at UIUC presents TILT (Training-Free Intrinsic Reward Tilted Sampling), a framework that addresses these failures without any fine-tuning or external reward models. The key insight is to interpret compositional failures as overlaps between joint and single-concept distributions, then define a reward that is intrinsic to the base model itself—favoring samples where all required concepts are simultaneously present.

TILT formulates this as a KL-constrained objective leading to a closed-form tilted target distribution that guides the diffusion sampling process. The interaction of concept distributions naturally gives rise to two distinct guidance strategies, which the authors combine in a hybrid approach for optimal performance. Experiments on the T2ICompBench benchmark show that TILT significantly improves compositional alignment (e.g., object attributes, relationships, counts) while maintaining or improving image quality compared to prior training-free and trained guidance methods. The method is simple, requires no external datasets or models, and can be applied to existing diffusion models at test time.