NextGAN teaches a critic where to guide fast image and video generation

Xinhao Lou and colleagues present NextGAN in an arXiv preprint submitted on November 4, 2026. Their goal is fast image and video generation by adapting pretrained models to produce outputs in one or a few generator evaluations. The paper reports experiments, not evidence that the method can train similarly large generators from scratch.

A critic that learns directions

A GAN has two competing components: a generator produces samples, and a critic gives the generator feedback. Conventional adversarial training teaches the critic to assign different values to reference and generated samples. But the generator is actually updated using the gradient of that critic: the direction in which its score changes. A critic that separates samples well does not necessarily provide the most useful update direction.

NextGAN trains that direction more explicitly. For a given condition, such as an image class or video prompt, it independently samples a reference example and a generated example. Frozen representation networks convert both into feature vectors. The method draws a point between those vectors and trains a scalar-valued critic so that its gradient at that point predicts the difference from the generated feature to the reference feature. The critic learns from many such random pairs.

The pairs are not fixed destinations for individual outputs, nor are they claimed to be an optimal transport matching. Their displacements are noisy, but regression pools them into a condition-dependent signal. The generator then follows the critic’s score uphill at its own generated features, with gradients passing through the frozen representation network. Critic and generator updates alternate; the critic and representers are discarded at inference.

The authors show that, for a fixed generator and averaged over uniformly sampled interpolation points, this gradient-regression objective is equivalent to an adversarial critic objective: a difference between reference and generated critic scores, penalized by gradient energy along the connecting paths. The two implementations need not produce identical updates on a finite minibatch. Their theoretical distribution-matching result also depends on assumptions, including a sufficiently expressive critic; it applies to the selected feature distributions, not automatically to the original images or videos.

What the experiments show

In a controlled two-dimensional experiment with eight Gaussian clusters, NextGAN reached an effective mode count of 7.995, compared with 2.770 for a matched conventional GAN after 30,000 generator updates. Both captured similar fractions of samples within the target modes, about 93%. This illustrates more balanced coverage in that setup, not a general guarantee against mode collapse.

For class-conditional ImageNet images at 256 × 256 resolution, the authors report one-step Fréchet Inception Distance, or FID, of 2.87 for a pixel-space JiT-B generator and 1.54 for a latent-space DiT-XL/1 using gradient regression. The adversarial form achieved 1.36 with DiT-XL/1. Lower FID is better on this distribution-level image metric. These are single generator evaluations without guidance; the paper notes that published baselines may use different evaluators, so comparisons are not perfectly controlled.

The choice of features mattered. In a JiT-B ablation after 20,000 updates, three frozen image representers together achieved FID 4.10, versus 5.55 for the best individual representer tested. That supports combining complementary representations in this experiment, rather than treating any single feature extractor as universally sufficient.

For text-to-video, the team adapted Wan2.1 models with 1.3 billion and 14 billion parameters using low-rank adapters. With those adapters disabled, the frozen backbone also supplied text-conditioned video features to the critic. One trained 1.3B model supported one-, two-, and four-step sampling. On the authors’ VBench evaluation at 832 × 480 resolution and 81 frames, its total scores were 80.80, 83.13, and 82.72 respectively; higher is better. The one- and two-step totals were the highest among the compared 1.3B methods at those budgets. At four steps, AnyFlow scored higher, at 83.54.

Scaling did not consistently improve the comparison. At 14B parameters, NextGAN scored 79.77, 82.73, and 82.02 at one, two, and four steps, below AnyFlow’s corresponding totals. VBench is a benchmark proxy, not a complete measure of visual quality, motion, or prompt adherence. The paper’s manually selected still frames are illustrative and cannot establish temporal quality on their own.

The cost behind the shortcut

The authors describe training as requiring no online teacher scores or trajectories, a narrower claim than independence from teachers or pretraining. Their large-scale experiments start from pretrained generators, and the video reference set consists of videos generated offline by Wan2.1-14B. How well the video method works with directly collected real-video data remains untested in the paper.

Feature matching can also miss differences that frozen encoders discard. The theory does not prove convergence for the finite neural networks and alternating optimization used in practice; moreover, the critic is supervised between endpoints while the generator queries it at generated endpoints. NextGAN offers a concrete way to shape a GAN’s training signal, with encouraging results on specific pretrained backbones and benchmarks, rather than a demonstrated universal recipe for fast generation. The source is version 1 of the arXiv preprint, arXiv:2611.06270v1.