Four-step distillation for diffusion models. Every frame on this page is four network evaluations or fewer.

Scroll

Meet PDMD

State-of-the-art step distillation

▶

1:13 · sound on

Music: “Goliath” by Scott Buckley, licensed CC BY 4.0.

TL;DR

Projected Distribution
Matching Distillation

PDMD improves the diffusion distillation with a one-line projection that filters student's training error. Nothing else changes: no auxiliary loss, no additional network, no extra model pass.

# r = x0_critic - x0_studentd = d - (d * r).sum() / r.pow(2).sum() * r

Zimo Wang^1,2, Junkun Yuan², Angtian Wang², Haotian Yang², Canyu Zhang², Siyuan Yuan², Xingchang Huang², Bo Liu², Yizhi Wang², Yiding Yang², Chongyang Ma², Gordon Guocheng Qian^2,†

¹UC San Diego²ByteDance^†Project lead

Paper arXiv

Code GitHub

Weights Hugging Face

The DMD update as a vector, and the PDMD update keeping only the component perpendicular to the residual. — DMD moves the student endpoint along −d. PDMD projects out the component of d parallel to the student–critic endpoint residual r, an estimate of critic error, to suppress critic error. The two frames decode a real update target during training — the PDMD target contains fewer artifacts.

DMD moves the student endpoint along −d. PDMD projects out the component of d parallel to the student–critic endpoint residual r, an estimate of critic error, to suppress critic error. The two frames decode a real update target during training — the PDMD target contains fewer artifacts.

Gallery

MiniMax-H3-33B distilled with PDMD. 1344×768, 345 frames at 24 fps, four network evaluations each.

Gallery · 2 NFE

MiniMax-H3-33B distilled with PDMD, down to two network evaluations. 1344×768, 345 frames at 24 fps, seed 42.

Check qualitative comparisons side by side →

Benchmark results

		Video metrics				Audio metrics
Method	NFE	Total	Quality	Dynamic	Semantic	PQ	CE	CU	IS
MiniMax-H3-33B, teacher	50	82.41	82.22	66.67	83.20	6.567	4.188	6.213	5.15
MiniMax-H3-33B	4	79.48	79.54	44.44	79.23	6.056	3.167	5.580	3.35
H3 Turbo LoRA	4	81.57	81.29	54.78	82.69	6.406	3.917	6.015	4.52
DMD2^†	4	82.27	82.51	59.95	81.34	6.305	3.434	5.871	4.06
DMD^†	4	82.76	82.68	61.76	83.05	6.381	3.801	5.931	4.79
rCM^†	4	81.18	80.98	58.40	81.95	6.063	3.711	5.446	3.69
AnyFlow^†	4	81.97	81.82	64.60	82.60	6.092	3.544	5.591	4.35
PDMD	4	83.17	83.25	71.83	82.86	6.530	4.062	6.180	4.98

387 prompts at 544p in mixed aspect ratios, seed 42 for every method. † marks our reimplementations, trained under the shared protocol; each reports its best checkpoint. Orange marks a column PDMD wins outright against every other four-step row.

Method	NFE	Total	Quality	Dynamic	Semantic
Wan2.1-T2V-1.3B, teacher	50×2	83.06	85.02	68.61	75.23
Wan2.1-T2V-1.3B	4×2	68.54	73.60	18.61	48.30
rCM	4	83.13	85.14	78.06	75.10
AnyFlow	4	83.54	85.28	59.44	76.57
DMD^†	4	82.70	85.06	86.39	73.24
DMD2^†	4	83.44	85.84	76.67	73.84
PDMD	4	83.73	85.89	89.72	75.09

944 AnyFlow-augmented VBench prompts at 480p, five seeds each. † marks our reimplementations, trained under the shared protocol; each reports its best checkpoint. Orange marks a column PDMD wins outright against every other four-step row. DMD^† and PDMD differ only by the projection.

User study

PDMD vs.

PDMD Tie Net (win − loss)

Bars are shares of all judgments in percent; hover a segment for the exact value. † marks our reimplementations.

← Back

Meet PDMD

Projected Distribution
Matching Distillation

Gallery

Gallery · 2 NFE

Read more

Qualitative Comparisons

Meet PDMD

Projected DistributionMatching Distillation

Gallery

Gallery · 2 NFE

Read more

Qualitative Comparisons

Projected Distribution
Matching Distillation