Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers
Open paper
Benchmark
Model
Rank
Results
text-to-image-generation-on-geneval
Qwen-Image-2512 with training-free head pruning (rank by image-to-semantic attention m_S, K=216)
#125
Overall: 0.758
Rank counts only results with a code link.