Get Started
Home
Topics
Search
Library
Research questionHow can attention heads be pruned in text-to-image diffusion transformers without losing prompt-specific object identity?During denoising, semantic information may be maintained by structural template tokens and image-to-text interactions rather than by the prompt tokens that initially encode it. This makes it difficult to identify redundant attention computation without disrupting object identity.
Diffusion Models
Image Generation
Inference Optimization
Mechanistic Interpretability
Multimodal Models
Latest papersRecent research connected to this question, newest first.Text Template Tokens Are Implicit Semantic Registers in Diffusion TransformersThe evidence concerns causal interventions and attention decomposition across token spans, heads, and layers in text-to-image diffusion transformers. It reports that structural template tokens become attention sinks that maintain object identity, while pruning heads that attend most strongly to prompt tokens removes 20% of attention FLOPs with a 1.4-point GenEval decrease.research paper · Jul 21, 2026
Related questions
How can high-resolution diffusion Transformers prune tokens without sacrificing image fidelity or predictable compute?How can attention-head contributions be measured in prompt-injection classifiers across circuit and output scales?How can low-rank compression preserve text-to-image quality in large diffusion transformers?How can preference optimization align whole-image preferences with token-specific content across spatial and denoising coordinates?