NitroFusion2 distills diffusion transformers to 1-2 steps via text-stream adversarial loss
SketchX · hf · 2026-11-05
Researchers at SketchX, University of Surrey, introduce NitroFusion2, which moves adversarial distillation from image features to the text stream of multimodal diffusion transformers (MMDiTs).
- Text-Stream Adversarial Distillation (TSAD) exploits intermediate text-token states as the adversarial representation, since joint attention means the text stream has already read the image
- Aspect-Cued Discrimination (ACD) conditions discriminators on aspects like typography, objects, or style to steer what adversarial training emphasizes
- Few-step distillation drops from 4-8 steps to 1-2: one-step generation yields sharp rendered text and fine details, and two steps stay competitive on benchmarks
More from Multimodal
- AI video demo of festival crowd movement goes viral on Reddit — shahryar2025 · 2026-11-05
- UniRestorer++ scales universal image restoration with multi-granularity degradation modeling, beating dense and MoE baselines — Jingbo Lin · 2026-11-05
- GaussianRenderFormer: Hebrew U team renders 3D Gaussian splats with a pure transformer, no rasterizer, hitting 38.08 dB — HUJI-IL · 2026-11-05
- Blueprint of a memory: visualizing an AI model's engram as an annotated neuron — RileyRalmuto · 2026-11-05
- Remiqora Video 0.3: open-source studio turns a line of text or a song into a film on 16GB RTX — inikolax · 2026-11-05
- Testing image 2.6's GIF feature: prompt an anime character to nod and close her eyes — RumbleversePlayer · 2026-11-05