Adversarial training enables one- and few-step image and video generation, but conventional adversarial losses supervise critic values without directly specifying the gradients used to update the generator. We present NEXTGAN, which trains a scalar critic by regressing its input gradient onto displacements between independently paired reference and generated features at interpolated locations in frozen representation spaces. For a fixed generator, this regression is equivalent to an adversarial objective with a gradient-energy penalty, while the generator retains the standard potential-based GAN update. The same objective adapts pre-trained image and video generators with up to 14B parameters, without auxiliary losses, online teacher scores, or trajectories. On ImageNet 256×256, NEXTGAN achieves one-step FIDs of 2.87 with pixel-space JiT-B and 1.54 with latent-space DiT-XL/1; its adversarial form further reaches 1.36. For text-to-video generation, a single trained Wan2.1-1.3B model supports one-, two-, and four-step sampling, achieving the highest VBench total scores among the compared 1.3B methods at one and two steps. At two steps, NEXTGAN scores 83.13, exceeding rCM and AnyFlow by 1.25 and 0.47 points, respectively. Code is available at https://github.com/LINs-lab/NextGAN.git.
Figures & tables
Figure 1: Qualitative samples from NextGAN . Left: 1-NFE class-conditional samples on ImageNet 256×256 . Right: four frames from each of four generated videos.
Figure 2: How NextGAN learns to guide the generator. xp and xr are generated and real features; xα lies between them. First, the critic loss LD trains ∇xuϕ to predict their displacement, with endpoints fixed ( sg : stop gradient). Then, the generator loss LG moves generated features uphill in the fixed potential uϕ . The surface depicts −uϕ , so this motion appears downhill. These two updates alternate throughout training.
Figure 3: Mode coverage during training. Rows: NextGAN and GAN. Columns: samples from fixed latent probes at selected iterations and the target distribution. The curves below each row report mode imbalance, 8−Neff ; lower is better.
Method
Space
Backbone
Params
NFE
FID ↓
IS ↑
Consistency / flow-map distillation
Shortcut ( Frans et al., 2025 )
latent
DiT-XL/2
676M
1
10.60
–
PaGoDA ( Kim et al., 2024 )
pixel
ADM
301M
1
1.56
259.61
Diffusion / flow matching
MeanFlow ( Geng et al., 2025 )
latent
XL/2
676M
1
3.43
–
ADM ( Dhariwal & Nichol, 2021 )
pixel
ADM-G + ADM-U
866M
250+250
3.94
215.8
Table 1: Class-conditional ImageNet-256 results. Baselines are quoted from their original papers and may use different evaluators. NextGAN Params denote nominal backbone sizes (implementation counts: App. A ). NFE counts sequential generator evaluations, including guidance and autoregressive inner steps where applicable.
Model
Online teacher
NFE
Evaluation Scores ↑
Quality
Semantic
Total
1.3B parameters
Wan2.1 T2V ( Wan Team, 2025 )
–
50×2
82.91
80.22
82.37
rCM ( Zheng et al., 2025 )
yes
1
81.76
76.34
80.68
AnyFlow ( Gu et al., 2026 )
yes
1
78.74
75.63
78.12
Wan2.1 + DMD2 ( Yin et al., 2024 )
yes
1
80.30
71.85
78.61
Table 2: Text-to-video generation on VBench at 832×480×81 . Results are grouped by model size and NFE. “Online teacher” indicates whether training requires online teacher scores or trajectories. Higher is better for all three scores.
Figure 4: Video generation with GAN, WGAN, TBSM (GAN), and NextGAN . Left: NextGAN achieves the highest quality, semantic, and total VBench scores (best in bold). Right: NextGAN shows clearer subject identity, more coherent object structure, and more accurate text generation.
Figure 5: Six manually selected prompts (columns) with manually selected frames at matching indices across methods. Rows: rCM, AnyFlow, and NextGAN (all 1.3B).
Figure 6: Intra-prompt sample diversity under four-step sampling (Wan2.1-1.3B). Left: Wan2.1 + DMD2; right: NextGAN . Rows show fish, flowers, and man; each panel shows five seeds per prompt, with matched seed indices across methods. Corresponding videos from both methods are shown at the same frame index.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Training objective and feature space. (a) JiT-B training curves through 50K updates (logarithmic FID-50K axis). (b) Frozen representer ablations on JiT-B at 20K updates. Lower is better.
Generative Adversarial Networks (GANs) train competing generators and discriminators to synthesize data, translate images, and model complex distributions.
We present AAD-1, an Asymmetric Adversarial Distillation framework for One-step autoregressive image-to-video generation. State-of-the-art methods adopt adversarial distillation but suffer from motion collapse and training instability, resulting in static videos. AAD-1 addresses these challenges through two key designs in architecture and training strategy. Our key architectural insight is to break the symmetry between generator and discriminator. While the generator remains causal to preserve autoregressive sampling capability, the discriminator attends bidirectionally over the full spatiotemporal context and produces a single holistic realism score for the entire video sequence. This asymmetric design enables the discriminator to effectively detect global temporal failures and long-range drift that cause motion collapse in autoregressive generation. To stabilize training, we introduce a phased strategy that first uses distribution matching to bootstrap a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. Extensive experiments on VBench demonstrate that AAD-1 achieves state-of-the-art performance in one-step autoregressive video generation.
Haobo Li, Yanhong Zeng, Yunhong Lu +6
AutoLab, SAI, SJTU · Ant Group · Department of Automation, Tsinghua University +2
We propose continuous adversarial flow models, a type of continuous-time flow model trained with an adversarial objective. Unlike flow matching, which uses a fixed mean-squared-error criterion, our approach introduces a learned discriminator to guide training. This change in objective induces a different generalized distribution, which empirically produces samples that are better aligned with the target data distribution. Our method is primarily proposed for post-training existing flow-matching models, although it can also train models from scratch. On the ImageNet 256px generation task, our post-training substantially improves the guidance-free FID of latent-space SiT from 8.26 to 3.63 and of pixel-space JiT from 7.17 to 3.57. It also improves guided generation, reducing FID from 2.06 to 1.53 for SiT and from 1.86 to 1.80 for JiT. We further evaluate our approach on text-to-image generation, where it achieves improved results on both the GenEval and DPG benchmarks.
Few-step autoregressive video generation enables low-latency streaming, yet prevailing training pipelines separate causal adaptation from distribution matching. Recent representation-based text-to-video methods retain this separation by starting with an ODE initialization stage, even when the target distribution is defined directly by reference videos. We introduce \method, the first single-stage framework that uses reference videos both as paired examples of how a history continues and as a target population for self-generated rollouts. Specifically, Direct Forcing is trained with two complementary signals: (1) Causal Continuation Alignment learns paired spatiotemporal correspondence under reference histories; and (2) Rollout Distribution Matching employs maximum mean discrepancy to align the representation distributions of self-generated and reference videos. To coordinate these signals, we introduce Dynamic Gradient Balancing, which adapts their relative gradient contributions during training. Experiments with Wan2.1-1.3B show that Direct Forcing learns a four-step autoregressive generator in 250 updates without a separate ODE initialization stage or online diffusion-score supervision, achieving 84.63 VBench Total and 86.01 Quality.