2026-11-05 · America/Los_Angeles · 论文 · #20

NitroFusion2: Bringing Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers

Adversarial diffusion distillation has conventionally relied on image features for discrimination. We revisit this choice for multimodal diffusion transformers (MMDiTs), whose text stream becomes image-dependent through joint attention. We introduce Text-Stream Adversarial Distillation (TSAD), which exploits intermediate text-token states as a new adversarial representation. Because the text stream has already read the image while remaining tied to the text condition, changing that condition changes what information it extracts from the image. Building on this property, we introduce…

热度 40.7 / 100；排名与评分保留该期记录。