Paper Overview
- Field: Computer Vision
- Authors: Xin Lin, Zhifei Zhang, Yuqian Zhou, Haitian Zheng, Shaoteng Liu, Lehan Yang, Zhe Lin, Ming-Hsuan Yang, Truong Nguyen
- Published: 2026-09-29
- arXiv: 2609.38156
- Teacher-matching side: Diagnostics show that low-noise RGB matching is dominated by a local-texture cue, motivating a fixed high-noise matching band.
- Real-data side: Native clean-RGB outputs allow guidance from an external visual representation without traversing a decoder or sharing a heavy fake-score critic. DINO-Adv removes this critic from the adversarial gradient path and supplies local parametric patch guidance.
- AF-Loss: A parameter-free auxiliary semantic distribution-field objective designed for text-to-image DMD. It operates on disjoint rolling real and generated support sets in a shared DINOv2 space while preserving prompt-conditioned teacher supervision. AF-Loss adds no learnable parameters or inference-time computation.
Summary
Distribution matching distillation (DMD) provides a general framework for few-step diffusion generation, but its modern text-to-image instantiations have been developed primarily around latent diffusion, overlooking the key properties and design opportunities of native RGB space. DMA² revisits two DMD interfaces for pixel-space teachers:
Results
Together, these designs constitute DMA². On DPG-Bench, GenEval, VQAScore, and COCO30K, the four-step DMA² student outperforms the 25-step teacher model and evaluated few-step distillation methods.
---
*Auto-collected on 2026-10-01.*