Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers
For researchers working on generative models, this work provides a causal understanding of internal mechanisms in DiTs, enabling efficient pruning without significant performance loss.
The paper introduces a causal interpretability framework for diffusion transformers (DiTs) and discovers that structural template tokens act as implicit semantic registers, maintaining object identity during generation. By pruning heads that attend most to prompt tokens, they achieve a 20% reduction in attention FLOPs with only a 1.4-point drop on GenEval.
Text-to-image diffusion transformers (DiTs) jointly process text and image tokens, yet their internal computation during denoising remains poorly understood. We introduce a causal interpretability framework for modern large-scale DiTs that combines attention decomposition with targeted interventions across token spans, heads, and layers. Using it to separate prompt-content tokens from structural template tokens, we find that the structural tokens carry little prompt-specific information at the encoder output. Yet surprisingly, they emerge as dominant image-to-text attention sinks and causally maintain object identity inside the DiT, acting as implicit semantic registers. We show that they acquire this identity indirectly, with prompt semantics first injected into the image latents and then read back into the template tokens rather than transferred directly from the prompt tokens. Inspired by the above findings, we design a training-free pruning rule for DiTs. Heads that attend most strongly to prompt tokens are dispensable, and pruning them removes $20\%$ of attention FLOPs with only a $1.4$-point drop on GenEval. We further reveal how generative computation in DiTs is organized across heads and depth, separating semantic routing from visual synthesis and progressing from identity formation to propagation and refinement. Our work not only reveals that the tokens encoding semantics at input need not be those that maintain it during generation, but also provides a causal view of internal mechanisms in DiTs.