●builderYou can now implement parallel token updates and bidirectional attention without full-model fine-tuning.
●researcherThis demonstrates a highly efficient method for inducing diffusion generation properties in existing transformer architectures.