Dar-Yen Chen and Yi-Zhe Song · One-step diffusion distillation

NitroFusion2

NitroFusion2 is a research paper on adversarial diffusion distillation for multimodal diffusion transformers (MMDiTs), by Dar-Yen Chen and Yi-Zhe Song. It introduces Text-Stream Adversarial Distillation (TSAD), which places the adversarial loss on intermediate text-token states as well as on image features. In an MMDiT the text tokens have already read the image through joint attention, so they hold a summary of the generated sample that stays tied to the text condition. It also adds Aspect-Cued Discrimination (ACD), where changing the cue (typography, objects, action, style or the full caption) steers what a single frozen MMDiT discriminator emphasises. The authors report that few-step distillation drops from the usual 4-8 steps to 1-2 steps at comparable quality, with sharp rendered text and fine detail at one step and competitive results on the evaluated benchmarks at two steps. The text does not give benchmark figures, a licence or a code release; the paper is listed as arXiv:2611.05648.