Few-step autoregressive video generation

Enhancing Autoregressive Video Generation
via Representation Adversarial Distillation

Radian complements on-policy distribution matching with real-data supervision in a frozen visual representation space, without changing inference cost.

Fangyu Lin^1,2 Xingtong Ge^1,2 Lunjie Zhu^1,2 Yi Zhang^2,‡ Zhening Liu¹ Tianhang Wang³ Mengfei Li^1,2 Yumeng Zhang¹ Guanglu Song² Yu Liu² Jun Zhang^1,†

¹ The Hong Kong University of Science and Technology ² Vivix Group Limited ³ Zhejiang University

‡ Project Lead · † Corresponding Author

Read the paper ↗ Watch comparisons ↓ Code coming soon

4 NFE · Bicycle

1† NFE · Panda dining

4NFE · Radian

Intro

Real-data supervision, where visual quality is represented.

Few-step autoregressive video generation enables efficient streaming synthesis, but early errors are reused as context and can accumulate into detail loss, structural drift, and unstable motion. Existing distribution matching distillation primarily aligns student and teacher distributions in diffusion latent space, without directly supervising the perceptual quality of decoded videos.

Radian sparsely decodes frames from on-policy student rollouts, maps them into a frozen visual foundation model, and trains lightweight discriminator heads against real videos. DMD anchors the student to the teacher distribution, while representation-space adversarial supervision supplies complementary perceptual and semantic signals. The feature encoder and discriminator are removed after training, so the generator architecture and sampling budget remain unchanged.

Decoded-output supervision

Discriminate generated and real samples after RGB decoding, making blur, texture defects, and VAE artifacts directly observable.

Complementary signal with no inference overhead

Keep DMD as the distribution anchor, introduce perceptual directions during training, then discard the feature encoder and discriminator at inference.

Broad feature-space exploration

Systematically compare image, video, joint image–video, and diffusion-internal representations under a controlled adversarial distillation setup.

Method

Conceptual comparison among DMD, DiT-GAN and Radian, illustrating complementary representation-space supervision. — **Complementary supervision.** DMD anchors the student to the teacher distribution; Radian contributes a real-data perceptual signal that corrects degraded modes in decoded output space.

A three-stage recipe for representation adversarial distillation.

Radian begins from a causal ODE initialization, calibrates DMD and the discriminator, then jointly optimizes the student with on-policy DMD and a scheduled representation-space adversarial objective.

Joint objective

ℒ_Radian(t) = λ_DMD ℒ_DMD + g_t λ_adv ℒ^Φ_adv

ℒ_DMD preserves the teacher-derived generative prior, while the scheduled ℒ^Φ_adv term reshapes outputs toward perceptually realistic and semantically coherent real-data modes.

The three-stage Radian training pipeline, including DMD score models and image/video representation adversarial supervision. — **Training pipeline.** Sparse decoded frames are processed by a frozen image or video foundation model. Lightweight heads distinguish generated features from real-data features; only the causal generator is retained at inference.

Lightweight multi-level discriminator heads attached to frozen visual foundation model features. — **Lightweight discriminator.** Independent heads convert multi-level frozen features into dense spatial realism logits.

Multi-level feature discrimination

Dense evidence at every selected feature level.

For each sparsely decoded frame, the frozen encoder supplies local tokens augmented with its global readout. A small trainable head then scores every spatial location; logits are averaged over time, levels, and tokens.

Frozen representation

F̃_ℓ(y_m) = Φ_ℓ(𝒫(y_m)) + 1 r^⊤_ℓ

Dense discriminator logits

d_m,ℓ(y_m) = h_ψ,ℓ(F̃_ℓ(y_m)) ∈ ℝ^N_ℓ

Real–generated hinge objective

ℒ_D = 𝔼_x[(1 − d(x))₊] + 𝔼_x̂[(1 + d(x̂))₊]

Training only. The VAE decoder, frozen VFM, and discriminator heads are removed after training, adding no inference-time parameters or NFE.

Quantitative Results

One method, three autoregressive regimes.

All tables below are redrawn from the paper's unified local evaluation. Darker purple denotes stronger column-wise performance.

Four-step chunk-wise generation15 shared seeds · 944 prompts per seed

4 NFE

Method	NFE	VBench ↑									VideoAlign ↑
Method	NFE	Total	Quality	Semantic	Subj. Cons.	Dynamic	Aesthetic	Imaging	Object	Color	VQ	MQ	TA	Total
Self Forcing	4	0.8393	0.8471	0.8082	0.9378	0.6972	0.6745	0.6995	0.9440	0.8680	0.0695	0.0616	0.3683	0.4995
Causal Forcing	4	0.8400	0.8495	0.8019	0.9310	0.8694	0.6766	0.6993	0.9544	0.8311	0.0072	-0.0951	0.3542	0.2663
Salt	4	0.8434	0.8548	0.7982	0.9383	0.7926	0.6797	0.7034	0.9494	0.8291	0.0472	-0.0310	0.3774	0.3936
DiT-GAN	4	0.8411	0.8483	0.8122	0.9492	0.5954	0.6832	0.6984	0.9563	0.8713	-0.0242	0.0855	0.3715	0.4327
Radian (CD init)	4	0.8439	0.8550	0.7993	0.9380	0.7972	0.6792	0.7037	0.9508	0.8300	0.0521	-0.0296	0.3754	0.3979
Radian	4	0.8444	0.8541	0.8054	0.9504	0.8056	0.6801	0.7148	0.9570	0.8722	0.2134	0.2094	0.3805	0.8033

One-step frame-wise generation5 shared seeds · first-frame enhancement

1† NFE

Method	NFE	VBench ↑							VideoAlign ↑
Method	NFE	Total	Quality	Semantic	Object	App. Style	Temp. Style	Overall Cons.	TA	Total
Causal Forcing++	1†	0.8383	0.8473	0.8025	0.9473	0.7248	0.6908	0.7255	0.3115	-0.1813
One Forcing	1†	0.8415	0.8533	0.7943	0.9568	0.7433	0.6907	0.7178	0.3168	0.0448
Radian	1†	0.8417	0.8511	0.8040	0.9663	0.7546	0.7034	0.7267	0.4105	0.2563

VBench-Long957 frames · 832×480 · 16 FPS

≈ 60 sec

Method	NFE	Total ↑	Quality ↑	Semantic ↑	Subject Cons. ↑	Temp. Flicker ↑	Motion Smooth. ↑	Dynamic ↑	Imaging ↑
Rolling Forcing	5	0.7805	0.8191	0.6260	0.9753	0.9871	0.9842	0.4532	0.7075
Radian	4	0.8041	0.8511	0.6160	0.9790	0.9882	0.9864	0.6741	0.7194

Ablation Study

Representation space, encoder scale, and feature depth.

Controlled comparisons under the same 4-NFE chunk-wise training and inference setup. Darker purple denotes stronger column-wise performance.

Adversarial feature spaces15 shared seeds · identical generator and inference procedure

Table 2

Feature source	Feature model	Total ↑	Quality ↑	Semantic ↑	Multi-Obj. ↑	Color ↑	Consist. ↑	Dynamic ↑	Imaging ↑	VQ ↑	MQ ↑	TA ↑	VA Total ↑
Image VFM	DINOv2	0.8444	0.8541	0.8054	0.8343	0.8722	0.2662	0.8056	0.7148	0.2134	0.2094	0.3805	0.8033
Image VFM	DINOv3	0.8401	0.8498	0.8012	0.8528	0.8729	0.2649	0.7685	0.7104	0.1462	0.2937	0.3607	0.8006
Image VFM	SigLIP2	0.8391	0.8475	0.8052	0.8384	0.8662	0.2680	0.6343	0.6992	0.1156	0.1320	0.3751	0.6226
Video VFM	V-JEPA 2.1	0.8360	0.8430	0.8078	0.8604	0.8627	0.2676	0.5991	0.7042	0.0498	0.1275	0.4119	0.5892
Video VFM	VideoMAE	0.8372	0.8436	0.8115	0.8417	0.8787	0.2676	0.5213	0.7026	0.1000	0.1452	0.3436	0.5888
Image + Video VFM	DINOv2 + VideoMAE	0.8434	0.8525	0.8070	0.8652	0.8545	0.2679	0.6593	0.7074	0.0831	0.0575	0.3079	0.4485
Generative	DiT-GAN	0.8411	0.8483	0.8122	0.8432	0.8713	0.2674	0.5954	0.6984	-0.0242	0.0855	0.3715	0.4327

DINOv2 encoder size15 shared seeds · frozen visual encoder

Table 3

Size	VBench Total ↑	Quality ↑	Semantic ↑	VA Total ↑
S	0.8444	0.8541	0.8054	0.8033
B	0.8322	0.8404	0.7993	0.8910
L	0.8301	0.8366	0.8042	0.3686
G	0.8387	0.8475	0.8035	0.3257

Number of DINOv2 feature layersDINOv2-S · 15 shared seeds

Table 4

# Layers	VBench Total ↑	Quality ↑	Semantic ↑	VA Total ↑
1	0.8182	0.8206	0.8085	0.6747
2	0.8283	0.8389	0.7859	0.6586
4	0.8444	0.8541	0.8054	0.8033
8	0.8381	0.8469	0.8026	0.7228

Qualitative Results

Watch every method on the same prompt.

Choose a generation regime and case, then use the shared controller to start every method on the same frame. Videos are muted for consistent playback.

More Visualization

Radian beyond the paper selection.

Additional prompts across natural scenes, stylized actions, and difficult human–object interactions.

Drag or shift-scrollVideos play when centered

Citation

Radian

@misc{radian,
  title  = {Enhancing Autoregressive Video Generation via Representation Adversarial Distillation},
  author = {Lin, Fangyu and Ge, Xingtong and Zhu, Lunjie and Zhang, Yi and Liu, Zhening and Wang, Tianhang and Li, Mengfei and Zhang, Yumeng and Song, Guanglu and Liu, Yu and Zhang, Jun},
  year   = {2026},
  note   = {Manuscript}
}

Enhancing Autoregressive Video Generationvia Representation Adversarial Distillation