Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers
Chat-template tokens end up carrying object identity inside diffusion transformers.
The paper says these tokens hold little prompt-specific information after text encoding, but become major image-to-text attention sinks during denoising. Its causal analysis finds they maintain semantic identity by drawing it from image latents after prompt meaning is injected early. The authors also report a training-free pruning rule that removes inert prompt-reading heads, cutting 20% of joint-attention FLOPs with a 1.4-point GenEval accuracy cost. HF Daily Papers' note
The paper says these tokens hold little prompt-specific information after text encoding, but become major image-to-text attention sinks during denoising. Its causal analysis finds they maintain semantic identity by drawing it from image latents after prompt meaning is injected early. The authors also report a training-free pruning rule that removes inert prompt-reading heads, cutting 20% of joint-attention FLOPs with a 1.4-point GenEval accuracy cost. HF Daily Papers' note
score 4