TILT improves compositional text-to-image generation with a model-intrinsic reward

Debottam Dutta · hf · 2026-07-29

TILT proposes a training-free way to improve compositional text-to-image generation

The paper, TILT: Improving Compositional Generation in Diffusion Models with a Model-Intrinsic Reward, targets a common failure mode in text-to-image models: handling complex prompts with multiple concepts.
It introduces a training-free test-time method that changes the sampling trajectory using a reward derived from the base model itself, so it does not need external supervision or a separate reward model.
The authors frame compositional failures as overlap between joint and single-concept distributions, then derive a KL-constrained objective with a closed-form tilted target distribution and guidance steps for diffusion sampling.
On T2ICompBench, the method improves compositional alignment while preserving image quality, and a hybrid guidance strategy performs best among the explored variants.

More from Multimodal