DMAD: Distribution Matching as Adversarial Distillation
for Fast Visual Generation

Trailer

Generated with our 4-step distilled MiniMax-H3 model.

1.04FID ↓

One-step images

ImageNet-64 × 64
With projected discriminator

14.47FID ↓

Text to image

SDXL · 4 steps
COCO-10K

85.15VBench ↑

Text to video

Wan2.1-T2V-14B
4 steps

84.6%

Human preference

MiniMax-H3 · vs. rCM
Overall preference, excluding ties

How DMAD learns

Two discriminator heads. One shared backbone.
Distribution matching through classification.

Training in motion
01 / 05
Animated explanation: noise passes through a few-step student; teacher, student and real samples are noised and classified by two heads on one backbone; gap-based weights guide the student update.
0.0 / 31.5 s

From noise to a moving world.

A few-step student turns noise into a generated sample.

StudentGeneratingDiscriminatorIdle

Results

Qualitative comparisons

The same prompts, side by side.

Teacher50 steps
DMAD Ours4 steps
0.0 / 5.2 s

Same prompt across methods.

View generation prompt

Quantitative results

From 64 × 64 images to 33B audio-video models.

EDM · ImageNet-64 × 64

Class-conditional generation · Lower FID is better

Selected EDM · ImageNet-64 × 64 benchmark results
MethodForward passesFID ↓
EDM teacher (SDE)5111.36
DMD211.51
DMD2 + longer training11.28
D2O-F11.16
SiD11.52
SiDA11.35
SiD²A11.11
DMAD Ours11.24
DMAD + projected discriminator Ours11.04

Human evaluation

MiniMax-H3 · 4 denoising steps

DMAD vs. DMD2
DMAD preference
Overall Preference79.1%

204 wins · 129 ties · 54 losses

Visual Quality84.2%

208 wins · 140 ties · 39 losses

Audio Quality87.9%

145 wins · 222 ties · 20 losses

Motion Quality60.3%

79 wins · 256 ties · 52 losses

Prompt Alignment54.3%

51 wins · 293 ties · 43 losses

DMAD vs. rCM
DMAD preference
Overall Preference84.6%

247 wins · 95 ties · 45 losses

Visual Quality91.3%

262 wins · 100 ties · 25 losses

Audio Quality86.0%

129 wins · 237 ties · 21 losses

Motion Quality73.8%

121 wins · 223 ties · 43 losses

Prompt Alignment61.1%

55 wins · 297 ties · 35 losses

Demo Video

Trailer, method overview and visual comparisons.

Watch the complete DMAD demonstration.Watch on YouTube

BibTeX

@misc{yu2026dmad,
  title  = {DMAD: Distribution Matching as Adversarial Distillation
            for Fast Visual Generation},
  author = {Zhengming Yu and Junkun Yuan and Haotian Yang and
            Gordon Guocheng Qian and Yizhi Wang and Angtian Wang and
            Yiding Yang and Bo Liu and Xin Li and Wenping Wang and
            Chongyang Ma},
  year   = {2026}
}

Generated sample

Generation prompt