TILT: Improving Compositional Generation in Diffusion Models with a Model-Intrinsic Reward
This work addresses the problem of compositional failures in text-to-image generation for users requiring precise multi-concept prompts, offering a training-free method that improves alignment while maintaining image quality.
TILT improves compositional generation in diffusion models by using a model-intrinsic reward to align sampling trajectories with complex prompts, achieving better compositional alignment on T2ICompBench without external supervision.
Recent advances in powerful text-to-image generation models have made it increasingly important to develop test-time methods that modify the sampling trajectory to produce images more faithful to complex compositional prompts. We present TILT, a training-free framework for compositional text-to-image generation via test-time reward alignment. We interpret compositional failures as overlap modes between joint and single-concept distributions, and define a reward that favors samples where all concepts are jointly present. This reward is intrinsic to the base model and does not require any external supervision or reward models. This yields a KL-constrained objective with a closed-form tilted target distribution and principled guiding steps for diffusion sampling. The interaction of concept distributions together with the above reward naturally leads to two different guidance strategies while a hybrid approach that balances their respective benefits produces stronger performance. Experiments on prompts from T2ICompBench show that our method improves compositional alignment while preserving image quality compared to previous baselines.