NitroFusion2: Bringing Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers
NitroFusion2 cuts image generation to one or two steps at quality the authors report as comparable to conventional four-to-eight-step distillation.
The method uses intermediate text-token states to judge generated images, exploiting a text stream that has already absorbed image information through joint attention. Aspect-Cued Discrimination directs that judgment toward properties such as typography, objects or style. The authors report sharp rendered text and fine detail at one step, with two-step generation remaining competitive across the evaluated benchmarks.
HF Daily Papers' note
The method uses intermediate text-token states to judge generated images, exploiting a text stream that has already absorbed image information through joint attention. Aspect-Cued Discrimination directs that judgment toward properties such as typography, objects or style. The authors report sharp rendered text and fine detail at one step, with two-step generation remaining competitive across the evaluated benchmarks.
HF Daily Papers' note
score 5