Accelerated Training-Free Character-Consistent Text-to-Image Generation Framework

We tackle character-consistent text-to-image (T2I) generation in a training-free setting. Shared attention with a subject mask enforces identity consistency across prompts but introduces substantial memory and latency overhead at inference. To mitigate this cost, we augment the stable diffusion baseline with two accelerations: token pruning, which removes redundant tokens, and adaptive guidance, which skips unnecessary computation during the diffusion process. Experiments on the ConsiStory+ benchmark show that our method outperforms recent state-of-the-art approaches in character-consistent T2I generation. Notably, it attains lower inference latency than both prior state-of-the-art methods and the SDXL baseline.