Prompt Reinjection: Alleviating Prompt Forgetting in Multimodal Diffusion Transformers

TL;DR

This paper identifies prompt forgetting in multimodal diffusion transformers during text-to-image generation and proposes a training-free prompt reinjection method to improve instruction-following and image quality.

Contribution

It introduces prompt reinjection, a novel training-free technique to mitigate prompt forgetting in MMDiTs, enhancing their performance.

Findings

Prompt reinjection improves instruction-following capabilities.

It yields better scores on preference, aesthetics, and quality metrics.

The effect is verified across multiple models and benchmarks.

Abstract

Multimodal Diffusion Transformers (MMDiTs) for text-to-image generation maintain separate text and image branches, with bidirectional information flow between text tokens and visual latents throughout denoising. In this setting, we observe a prompt forgetting phenomenon: the semantics of the prompt representation in the text branch is progressively forgotten as depth increases. We further verify this effect on three representative MMDiTs--SD3, SD3.5, and FLUX.1 by probing linguistic attributes of the representations over the layers in the text branch. Motivated by these findings, we introduce a training-free approach, prompt reinjection, which reinjects prompt representations from early layers into later layers to alleviate this forgetting. Experiments on GenEval, DPG, and T2I-CompBench++ show consistent gains in instruction-following capability, along with improvements on metrics capturing preference, aesthetics, and overall text--image generation quality.