Main Question
How should intermediate generation be steered so that the denoised samples at t=0 are drawn from the pure modes?
TILT: Model-Intrinsic Reward Alignment
for Compositional Diffusion
Abstract
Diffusion models often struggle with compositional prompts, producing samples in which some concepts dominate while others are missing or weakly represented. TILT is a training-free framework that poses eventual pure-mode sampling as a reward for intermediate-time alignment. The reward is intrinsic to the model, yields a tractable target under a variational approximation, and can be applied across modalities. Experiments on vision and audio benchmarks show improved compositional alignment while preserving sample quality compared with prior baselines.
Pure Modes
A pure mode is a region of the final clean-sample distribution that represents the full composition. Its samples express the requested concepts together, without one dominating or disappearing.
Model-Intrinsic Reward
TILT rewards intermediate states that lead to pure modes at the final sample. The reward comes from the diffusion model itself, contrasting joint and constituent conditions without an external reward model.
TL;DR: Instead of heuristically correcting intermediate diffusion states, TILT treats landing in a pure compositional mode at the final sample as a model-intrinsic reward and derives test-time guidance toward that target.
Text-to-Image
Qualitative comparisons across compositional prompts.
Prompt: a green school bus and a red bag
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: The soft yellow duckling swam next to the sleek black swan.
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: a metallic jewelry and a wooden spoon
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: an oval coffee table and a square end table
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: a photo of a pizza right of a banana
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: a photo of a person and a sink
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: a photo of a surfboard and a suitcase
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Prompt: a photo of a red train and a purple bear
SDXL
CFG++
CO3
DoS
SuperDiff
R2F
TILT (Ours)
Text-to-Audio
Listen to SCORE-CLAP and SCORE-TILT for the same prompt.
Prompt: A series of doors sliding open as gusts of wind blows and glass clanks
SCORE-CLAP
SCORE-TILT (Ours)
Prompt: Water running softly in the background followed by several shrill beeps and a man speaking
SCORE-CLAP
SCORE-TILT (Ours)
Prompt: Wind noise while a water vehicle is traveling across water, and a man talks
SCORE-CLAP
SCORE-TILT (Ours)
Prompt: A gun firing several times followed by a revolver chamber spinning and metal clanking as a man talks and a person grunts
SCORE-CLAP
SCORE-TILT (Ours)
Prompt: A man speaks as rain pitter-patters and thunder rumbles
SCORE-CLAP
SCORE-TILT (Ours)
Method and pure-mode intuition
Compositional diffusion models can drift toward regions where a joint prompt behaves too similarly to one constituent concept, causing concept dominance or omission. Prior correctors suppress these collisions locally at intermediate noise levels.
TILT instead defines the desired outcome in data space: a final clean sample should lie in a pure joint mode. The resulting reward is intrinsic to the pretrained model, contrasting joint-condition likelihood against constituent-condition likelihoods.
The reward is computed from the diffusion model itself rather than an external vision-language or preference model. Under the paper's variational approximation, it becomes a contrast between the model's joint and single-concept denoising behavior.
This makes the same principle usable without training a modality-specific reward model.