DMA$^2$: Pixel-space Distribution Matching with Adversarial and Anchor Losses

What it claims

The paper adapts distribution matching distillation (DMD) to a pixel-space teacher, DeCo, a frequency-decoupled RGB diffusion model. It changes three things against the DMD2 recipe:

• DM-Band. Teacher matching is restricted to a fixed high-noise range. The summary does not mention this part.
• DINO-Adv. The adversarial head moves off the 560M-parameter fake-score critic and onto a frozen 86M DINOv2 ViT-B/14 with 1.58M trainable head parameters. This cuts same-pipeline training step time from 1.82 s to 0.72 s, a 2.5× speedup.
• AF-Loss. A parameter-free attract/repel objective in DINOv2 CLS space.

The four-step student beats the 25-step DeCo teacher on all seven reported metrics. The authors say the recipe also transfers to a second pixel backbone, PixelGen-XXL (Appendix B).

How it works

Generator. A single generator produces a clean RGB image in K steps, using the schedule S4 = {0, .25, .5, .75, 1} or S1 = {0, 1}. There is no VAE at any point. Only the generator, the fake-score critic and the adversarial heads are trained. Everything auxiliary is dropped at inference.

DM-Band. The authors ran a classifier two-sample test on 1,536 matched real/generated pairs. Real and generated RGB images are separable mainly near clean, through a local high-frequency texture cue. The cue survives a high-pass filter and a 32×32 patch shuffle but weakens under a low-pass filter. DINOv2 features, by contrast, stay discriminative into high noise.

Matching levels are therefore sampled from t_m ~ U(.02, .35), where lower t means more noise in DeCo's convention. The previous range was U(.02, .98). The band applies only to the DMD branch. It is a single hand-set value with no schedule and no online estimation.

DINO-Adv. Real and generated images are each noised at an independently sampled level. Patch features from DINOv2 layers {2, 5, 8, 11} feed small heads (1×1 conv 768→512, GroupNorm, LeakyReLU, 1×1 conv →1). The heads are trained with a non-saturating softplus GAN loss.

AF-Loss. The generated image's normalized final CLS feature z is attracted toward a real-feature bank and repelled from a generated-feature bank. Each bank is a detached FIFO queue of 256 features. Kernel radii are {.02, .05, .2}.

The loss is ‖z − sg(z + V(z))‖², so gradient flows only through z into the generator. It shares the DINOv2 forward pass with DINO-Adv.

flowchart LR
  G[Generator Gθ] --> X[clean RGB x̂]
  X --> N[re-noise t_m in .02–.35]
  N --> T[frozen teacher]
  N --> C[fake-score critic]
  T & C --> DMD[L_DMD]
  X --> D[frozen DINOv2]
  R[real images] --> D
  D -->|patch feats| ADV[L_GAN heads]
  D -->|CLS + FIFO banks| AF[L_AF]
  DMD & ADV & AF --> L[L_G = DMD + λGAN·GAN + λAF·AF]
  L --> G

Evidence

Band ablation on DPG-Bench. On the pixel-space DeCo student, DPG-Bench rises from 82.55 to 82.96 when the matching range narrows from U(.02, .98) to U(.02, .35).

Latent negative control. Applying the same band to a latent SDXL-DMD2 model lowers DPG-Bench from 67.64 to 63.23. The authors take this to mean DM-Band is specific to native RGB, not a general DMD schedule.

Training cost. Per-step training time drops from 1.82 s to 0.72 s with DINO-Adv.

Overall results. The headline results against the teacher and the few-step baselines (Flash, TDM, Decoupled DMD, DMDR) cover DPG-Bench, GenEval, VQAScore and COCO30K. The tables themselves are not in the fetched text.

How to use it

The fetched text gives no code, weights, licence or commands. Qualitative samples are shown at 512×512.

Limits

• Band setting. The band does not transfer to latent teachers: it hurt SDXL by about 4.4 DPG points. If you use a different pixel teacher, re-check the U(.02, .35) cutoff, which was chosen from DeCo-specific diagnostics.
• Size of the band gain. The isolated benefit of DM-Band is modest, +0.41 DPG.
• Truncated page. The fetch ends partway through Section 3.2. The main results tables, the component ablations for DINO-Adv and AF-Loss, the λ weights, the training compute and any stated limitations were not available. Only the abstract-level claims back the "beats teacher on all metrics" statement here.