multimodalfactneutralBlocking attention between tokens at different noise levels in diffusion models does not degrade performance and can even improve itComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02508v1