Continuous Adversarial Flow Models

Abstract: 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.

• The paper presents a continuous adversarial flow model that refines finite-capacity generative processes by integrating a learned discriminator within the ODE dynamics.
• It employs a Jacobian-Vector Product-based discriminator to align the generated trajectories with the true data manifold, overcoming MSE limitations in flow matching.
• Empirical evaluations show notable improvements in FID scores for image and text-to-image synthesis, demonstrating the method’s effectiveness as a post-training refinement.

Continuous Adversarial Flow Models: Technical Synthesis and Implications

Introduction and Motivation

Continuous Adversarial Flow Models (CAFMs) present a new paradigm for training continuous-time generative flow models using adversarial objectives, diverging from the canonical flow matching (FM) approach that relies on fixed mean squared error (MSE) objectives. The authors identify the sub-optimality of $L_2$-based criteria in FM, which overlooks the structure of the underlying data manifold, leading to the generation of out-of-distribution (OOD) samples—an issue especially prominent in realistic image and video synthesis. CAFMs address this by integrating a learned discriminator into the continuous-time domain, pushing the generator to synthesize samples better aligned with realistic data distributions.

Methodology

Flow Matching and Its Limitations

Conventional FM trains a generator $G(x_t, t)$ to predict the velocity vector driving a state $x_t$ (interpolating between noise $z$ and data $x$) toward data under some probability flow, minimized under MSE loss. In the infinite-capacity regime, all consistent loss functions yield the target marginal velocity; however, under realistic, finite-capacity models, the specific criterion directly sculpts generalization. The Euclidean metric used in FM is agnostic to the nonlinear geometry, thus producing inferior generalization by not respecting the true data manifold, with generation quality only partially rescued by external guidance (e.g., classifier-free guidance), which itself distorts the sample distribution.

Adversarial Training in Continuous Time

CAFMs generalize Adversarial Flow Models (AFMs)—which are discrete-time—to continuous time, connecting adversarial learning and continuous normalizing flows (CNFs). The generator learns a continuous vector field by integrating an ODE from noise to data, but it is now trained not just to minimize mean velocity error but also to defeat a discriminator $D$ parameterized over $(x_t, t)$.

A distinctive aspect of CAFMs is the adversarial loss formulation in the derivative space via a Jacobian-Vector Product (JVP) of the discriminator:

$$D_{\textrm{jvp}}(x_t, t, v_t, T) = \frac{\partial D(x_t, t)}{\partial x_t} v_t + \frac{\partial D(x_t, t)}{\partial t} T$$

This enables discrimination between the ground truth and generated velocity fields at every continuous state along the flow, aligning the adversarial signal with the physics of the generative process. The optimization is cast as a minimax game, with the generator seeking velocity predictions that maximize the discriminator's assessment of realism, and the discriminator penalized (with centering and optional optimal transport regularizers) to prevent trivial solutions and null-space exploitation.

Figure 1: Visualization of the training dynamic. Top: learned $G(x_t, t)$ trajectories over the probability flow. Bottom: corresponding $-D(x_t, t)$ values at all $G(x_t, t)$0. $G(x_t, t)$1 is taken for more intuitive visualization as the generation process runs backward in time.

Pre-training Versus Post-training and Implementation

CAFMs can be used for end-to-end training but are computationally less efficient due to the extra discriminator and JVP/gradient evaluation. They are best suited as a post-training procedure to refine pretrained FM-based models, yielding a substantial gain in distribution alignment for minimal extra resource usage. The implementation exploits forward-mode autodiff and vmap for efficient batched JVPs, and transformer architectures with RMSNorm normalization for stability, supplanting LayerNorm as per empirical ablation.

Empirical Evaluation

ImageNet 256px Generation

CAFMs achieve pronounced improvements on both latent-space and pixel-space ImageNet generation. Notably, post-training a pretrained SiT-XL/2 with CAFM reduces the guidance-free FID from 8.26 to 3.63 and the best classifier-free guided (CFG) FID from 2.06 to 1.53. A similar pattern is observed for pixel-space JiT models, with FID dropping from 7.17 to 3.57 without guidance and from 1.86 to 1.80 with guidance. These gains are robust across hyperparameter sweeps and are consistent with improvements in IS, sFID, precision, and recall.

Figure 2: Generation without guidance. Our method yields better generalization.

Qualitative differences further highlight considerable improvements in fidelity and semantic alignment in non-guided samples. When compared with recent baselines employing similar architectures and distillation schemes, CAFM post-training advances the state-of-the-art among large, open models in both pixel and SD-VAE latent representations.

Text-to-Image Generation

CAFMs generalize beyond class-conditional generation. Post-training Z-Image (a 6B single-stream transformer) with CAFM, the approach advances GenEval overall scores (from 0.81 to 0.85 with guidance) and DPG scores (from 83.7 to 85.2) in text-to-image benchmarks. These results are consistent across guidance scales and prompt expansions.

Figure 3: A photo of a dog.

Continuous Adversarial Flow Training from Scratch

Direct training with CAFM is possible but converges slower than FM-only, which recommends the post-training regime for computational efficiency.

Figure 4: SiT-B/2 pre-training on ImageNet 256px generation. Although CAFM can be used for pre-training, it converges slower than FM under the same epochs, fitting our expectation that CAFM is more suitable for post-training.

Theoretical and Practical Implications

CAFMs ground adversarial feedback within the actual generative ODE, discriminating not against discrete samples but their infinitesimal dynamics. This formulation (i) respects the manifold structure and (ii) aligns the generator toward trajectories that the data distribution actually supports, as attested both qualitatively and via standard statistical measures.

Notably, CAFMs mitigate classical vanishing-gradient issues that often hamper adversarial training in high dimensions. The JVP-based discriminator provides persistent, informative gradients because it linearly projects local generator actions along physically meaningful directions. Empirical results show that gradient penalties and discriminator data augmentation can often be omitted, simplifying optimization.

Crucially, the approach is generalized with respect to the ground-truth probability flow—FM and CAFM have the same optimal solution—but CAFMs exploit finite model capacity and adversarial gradients to yield superior generalization. In the context of few-step generators, CAFMs offer an orthogonal route to guidance and distillation, further improving sample quality without inadvertently distorting the learned distribution, as occurs with traditional classifier-based guidance.

Speculation on Future Directions

CAFMs open multiple avenues for generative model research:

• Divergence and Loss Design: The theoretical underpinning accommodates alternative loss metrics and divergence formulations, including Riemannian or perceptually-motivated objectives, further aligning learned manifolds with human perception.
• Higher-Dimensional and Multimodal Domains: The scalability and stability of JVP-based discriminators could facilitate improved synthesis in high-resolution video, multimodal (vision-language-audio) flows, or conditional generation tasks.
• Manifold Regularization: The adversarial loss could be combined with explicit manifold-awareness or learned intrinsic geometries, leading to further theoretical advancements.
• Adversarial Post-Training in Other Architectures: The technique may be productively applied in settings beyond canonical transformers, such as graph flows or continuous-time autoregressive models.

Conclusion

Continuous Adversarial Flow Models introduce a highly effective mechanism for closing the gap between finite-capacity generative models and the true data-supporting manifold via adversarial training in the continuous-time domain. As a post-training strategy, CAFMs consistently improve sample fidelity, semantic alignment, and distribution matching without sacrificing computational viability, all while preserving the formal connection to the ground-truth flow. The methodology is poised to serve as a cornerstone for future adversarial generative modeling in both supervised and unsupervised large-scale synthesis.