image generation teams

Abstract: An effective visual tokenizer must preserve image fidelity while producing latents that downstream diffusion models can learn efficiently. Prior work improves latent diffusability through semantic alignment or joint tokenizer--diffusion training. Revisiting REPA-E, a representative approach, we find that removing its diffusion objective and dedicated generative blocks while retaining semantic prediction from corrupted latents yields comparable downstream generation performance under a matched setting. This observation motivates a general principle of robust prediction: generation-relevant information should remain recoverable from corrupted latents. We instantiate this principle with two complementary objectives. RobustAlign predicts pretrained semantic features from corrupted latents to preserve high-level structure, while RobustRecon reconstructs the clean image to retain fine-grained spatial and appearance information. Together, they form Robust Alignment and Reconstruction VAE (RoAR-VAE). Our analysis suggests that robust prediction improves diffusability by making useful information more accessible in high-variance latent directions, which diffusion models tend to learn earlier. Empirically, RoAR-VAE improves average generation benchmark performance by 2.19 points over the strongest baseline tokenizer under the same training budget, while improving small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE. These results demonstrate that RoAR-VAE achieves a favorable balance between reconstruction fidelity and downstream generative learnability.

Abstract: Few-step generative models produce high-quality samples in one or a few evaluations. However, steering them toward desired properties remains difficult, especially when only black-box reward evaluations are available. While one could apply approaches developed for iterative diffusion and flow-based models, such as direct RL fine-tuning or test-time scaling, we find that neither transfers well to the few-step regime: direct fine-tuning suffers from severe reward over-optimization, and test-time scaling sacrifices the efficiency advantage of few-step generation. We propose Noise Policy Optimization (NPO), which explicitly targets the alignment problem for few-step generative models. The key idea is to learn an amortized flow policy over the initial noise space of the generator. The trained policy transports Gaussian noise to a reward-tilted noise distribution, whose pushforward distribution through the frozen generator is exactly the reward-tilted output distribution. By keeping the generator frozen, we can effectively utilize the knowledge encoded in the pretrained model, yielding a better reward-quality trade-off than direct fine-tuning. The amortized policy moves the search cost into training, so inference adds only a lightweight noise sampling step on top of the pretrained model. To improve sample efficiency during training, we use ELBO-based likelihood estimation that requires only sampled endpoints rather than simulated trajectories. Across image, language, and protein structure generation, NPO consistently achieves a better reward-quality trade-off than direct fine-tuning, and matches test-time search methods with 16 to 549 times lower inference cost.

Abstract: Few-step distillation reduces the sampling cost of diffusion and flow models, but existing approaches often require auxiliary score models or specialized architectures. We show that simple on-policy velocity regression already provides a strong foundation for 4-NFE generation. To represent teacher velocity variation within each coarse transition, we introduce Pivot, a training-only residual proxy that separates local corrections from the student's interval-mean prediction. We further repel the student from weaker teacher predictions, providing a lightweight mechanism for avoiding undesirable modes. The resulting framework, Simplified Matching Distillation (SMD), remains entirely regression-based, requires neither online score estimation nor adversarial training, and leaves the deployed generator unchanged. We evaluate SMD on image generation and editing, 3D generation, video synthesis, and joint audio--video generation, scaling to 33 billion parameters. At 4-NFE, SMD outperforms strong few-step baselines and surpasses multi-step teachers on several benchmarks. It also supports cross-model, multi-teacher, and causal-video distillation.

Abstract: Verifier-guided inference-time scaling uses additional computation to score intermediate states and steer generation toward better outputs. However, imperfect intermediate verifiers can distort sampling, and repeated refinement alone does not ensure recovery of the desired reward-tilted distribution. We introduce Continuous-State Verifier-Guided Backtracking (CS-VGB), an inference-time scaling method for sampling from reward-tilted distributions with continuous generative models through repeated denoising and renoising. With correctly paired denoising and renoising transitions, the stationary distribution of CS-VGB at the clean checkpoint equals the desired reward-tilted target, even when intermediate verifiers are inaccurate. Under stated regularity conditions, CS-VGB converges exponentially to the target despite verifier error, and the expected number of updates needed to reach a clean sample grows only linearly with the number of checkpoints. In analytically tractable settings, we empirically demonstrate that CS-VGB recovers the desired reward-tilted distribution despite verifier error, whereas the compared verifier-based sampling methods fail to recover the target under the same verifier errors. We also instantiate the framework with diffusion models and one-step stochastic posterior maps. Across prompt-guided ImageNet generation, facial-attribute generation on CelebA, and multi-robot planning, recurrent CS-VGB search improves reward and task success as inference budgets grow, with image-generation rewards continuing beyond the observed plateaus of denoising-only baselines.

Abstract: This paper studies the backbone used by metrics for evaluating the quality of generated images. Although generative models have advanced rapidly, some outputs still exhibit unrealistic texture characteristics, and these differences are difficult to measure. Widely used evaluation metrics provide limited insight into texture realism, since they often rely on a backbone trained for another objective, such as image classification or image-text alignment. Such objectives typically encourage invariance to low-level variations that carry texture, so the resulting features primarily encode semantic and global structure. As a result, texture-specific discrepancies may not be explicitly captured. We propose TRIG, a metric directed at the texture realism of generated images. Its backbone is instead the discriminator of a generative adversarial network pre-trained on a texture dataset, which provides a learned prior for the local statistics of natural textures. A plausibility head is fine-tuned to identify local texture degradations, while preserving natural defocus and sensor noise as positive cases. On the features of this backbone, we compute a standard distributional distance between a reference set of real images and a set of generated ones. We apply TRIG to several generative models and observe closer agreement with human realness ratings than baseline metrics.

Abstract: Visual representations are largely developed for understanding, yet what makes them useful for generation remains unclear. We study this question in pixel diffusion transformers and uncover two key insights. First, better representation quality does not necessarily translate into greater generative utility, as externally aligned semantics can be progressively lost during pixel synthesis. Second, architectural separation with specialized pixel decoders mainly redistributes semantic and spatial modeling, without necessarily yielding representations better suited to generation. These insights suggest that a useful generative representation must be both properly formed and actively exploited. Motivated by these insights, we introduce Representation Guidance (RG), which directly forms and exploits internal representations while retaining a plain Transformer architecture. A low-dimensional bottleneck forces generation-relevant information into a compact representation, which is then explicitly used to guide decoder generation by modifying the flow-matching target according to its induced prediction offset. RG learns clearer semantic and spatial structure while scaling efficiently in both parameters and computation. On ImageNet 256x256, RG-L/16 with only 459M parameters outperforms substantially larger pixel diffusion transformers, while RG-XL/16 achieves an FID of 1.55. RG further reaches an FID of 1.74 at 512x512 and extends to text-to-image generation, achieving 0.89 on GenEval and 87.1 on DPG-Bench.

Abstract: Reward-guided inference-time alignment can improve joint audio-visual generation, but optimizing multiple alignment objectives requires many expensive model evaluations, making existing methods costly under limited compute. Particle-based search enables global exploration but incurs estimation error with a finite population, while Metropolis-Hastings (MH) can refine candidates locally but benefits from a strong initialization. We introduce \textbf{S}pend \textbf{W}here \textbf{I}t \textbf{M}atters (\modelname), a two-stage, training-free, reward-agnostic framework that combines particle search with MH refinement under a fixed compute budget. In the first stage, we adaptively allocate particles across denoising checkpoints using online variance estimates, minimizing estimation variance while concentrating compute where it is most useful. This search provides a promising candidate for MH refinement, which uses local noise-and-denoise proposals on completed outputs to further improve reward. By starting from a strong candidate, MH can recover additional reward without requiring a larger particle population or costly global exploration. We show that, at matched compute, combining particle search with MH refinement achieves higher rewards than either approach alone. On \textsc{JavisBench-Mini}, at closely matched compute, \modelname achieves mean relative gains across seven metrics of $30.7\%$ over BoN and $14.1\%$ over EvoSearch with JavisDiT++, and $9.2\%$ over BoN and $17.0\%$ over EvoSearch with LTX-2. Beyond joint audio-visual generation, we also demonstrate the benefits of adaptive particle allocation on SD1.5 image generation, improving AestheticScore by $2.2\%$ and ImageReward by $6.0\%$ over uniform allocation.

Abstract: Vision-language reward models are trained on annotated preference pairs, which are expensive to collect, can disagree, and are ranked by benchmark accuracy. We build one with no human preference label: a linear head on an intermediate layer of a frozen Qwen3-VL backbone, trained to score a web caption above a perturbed version of the same caption; labeled data enters only to select the layer. The layer is the decisive choice, worth 5.8 points over the last layer, and one recipe transferred without re-tuning beats or at least matches GPT-4o on three benchmarks at 32B, 6.6 points ahead on VL-RewardBench, from a few thousand parameters at about a thousandth of fine-tuning compute. At an equal pair budget, perturbations teach a frozen head 3 to 5 points more than curated RLAIF-V pairs. Because the backbone is frozen, the objective is convex, and the perturbation sampler specifies the reward, which the benchmark does not identify: two heads trained on disjoint halves of the same perturbations tie on the benchmark and differ by 6.4 points in hallucination judging. Retraining the head on the policy's own verifier-labeled candidates cuts caption hallucination by 5.1 CHAIR points and lifts image-generation pass rate by 5.4, while scoring lower on every reward benchmark. Together, an intermediate-layer probe, free same-image perturbations, and a verifier-driven bootstrap turn give a recipe for multimodal reward modeling that needs no annotator, costs a fraction of fine-tuning, and is specified by the comparisons it is trained on rather than by a benchmark rank.

Abstract: Optimal-transport couplings have been shown to reduce training variance in unconditional flow models, but their role in conditional generation remains unclear. A natural approach constructs separate couplings for each condition, but this is impractical for large or continuous conditioning spaces found in modern image foundation models. We introduce Global Transport (GT), a global class-agnostic optimal-transport coupling, computed without class labels. GT can associate different conditions with different regions of the source noise, and consequently worsens performance without guidance. However, when combined with classifier-free guidance (CFG), GT consistently improves generation across domains, model scales, and sampling budgets. This reversal suggests that couplings for conditional flows should be evaluated both empirically and theoretically under the guided flow used at inference, rather than on unguided generation. We evaluate GT over both discrete class and continuous text conditioned image generation across model scales, and investigate how coupling choice alters guided trajectories. These results identify coupling design in the guided flow setting as a simple training time axis to improve performance without modifying existing architectures, samplers, or guidance mechanisms.

Abstract: Discrete diffusion models and flow matching have emerged as powerful frameworks for generative modeling over discrete state spaces, yet efficient few-step generation remains a fundamental challenge. In this work, we introduce the Discrete Average Generator, a principled extension of MeanFlow to Continuous-Time Markov Chains (CTMCs). Analogously to how MeanFlow defines an average velocity field over a time interval in continuous spaces, we define an average generator as the normalized increment of the transition kernel over a time interval. We show that this average generator satisfies a self-consistency identity, which provides the foundation for our training objective. We further develop training strategies that align with the standard training paradigm of diffusion language models while keeping the resulting objective tractable. When projected onto per-coordinate marginals, the self-consistency identity admits a closed-form expression, enabling efficient training and inference. In Potts model simulations, our objective reduces the total variation distance of the $K$-step sampler by up to 67%. On OpenWebText, our method achieves the lowest generative perplexity among the evaluated methods for 8 to 64 sampling steps while enabling a $16\times$ acceleration, and achieves comparable performance to existing methods on ImageNet.

Abstract: Generative models often suffer from mode collapse and limited sample diversity. While prior works attempt to mitigate this by jointly generating a batch of samples and repelling their trajectories, these heuristics do not explicitly maximize the diversity of the resulting endpoints. We introduce JIVE, a training-free framework that enhances generative diversity by injecting velocity perturbations aligned with the leading right singular subspace of the generator's endpoint Jacobian. By leveraging this local geometric structure, JIVE provably maximizes endpoint diversity while preserving sample quality. To maintain practical efficiency, we compute these perturbation directions via matrix-free iterations rooted in classical numerical linear algebra, requiring only a small computational overhead. Across different benchmarks, JIVE boosts both pixel and feature-level diversity in few-step and one-step generation.