#01Sep 10, 2026
cs.CV
Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling
Meimingwei Li, Stefan Andreas Baumann, Felix Krause and 1 more
Visual Autoregressive Models (VAR) generate images through next-scale prediction, producing all tokens within each scale in parallel. We show that this parallel decoding constitutes a mean-field-style approximation that discards spatial dependencies among same-scale tokens, causing locally incoherent samples regardless of backbone capacity -- a limitation of the decoding rule. Addressing this limitation, we introduce the Logit Refiner, a lightweight autoregressive module that restores intra-scale dependencies by sequentially sampling tokens conditioned on frozen backbone features. Adding only ~10% parameters and less than 5% of the base model's training compute, it plugs into any pretrained VAR checkpoint without retraining. Controlled ablations isolate joint intra-scale sampling -- rather than additional capacity or training -- as the critical ingredient. Across backbones from 310M to 2B parameters on class-conditional ImageNet 256x256, the refiner consistently improves generation quality, enabling a 1.1B-parameter model to surpass one twice its size. The approach further generalizes to text-to-image generation, confirming that the mean-field bottleneck persists across VAR variants and is effectively alleviated by our method. Project page: https://compvis.github.io/logit-refiner/
#02Sep 10, 2026
cs.CV
Learning Interaction between Image and Layout Priors for Joint Image-Layout Generation in Design Templates
Shirong Yang, Bo Yang, Ying Cao
In this paper, we address the problem of graphic design template creation, which generates a background image and a layout of foreground elements over the background to form a harmonious composition from an input text. Prior work on graphic design generation mostly adopts a sequential paradigm, where design elements are generated sequentially. We argue that such a sequential scheme falls short of faithfully capturing the dependency between the background and layout (and thus the joint image-layout distribution), which limits the quality of generated design templates. To overcome this limitation, we propose a model, InterIL, which jointly generates the two modalities, background image and layout, in a single generative process. The novel design of our joint model connects the backbones of pretrained image and layout diffusion models with a learnable communication module to explicitly model bidirectional image-layout interaction. During training, the image and layout backbones are frozen to maintain and leverage the vast pretrained single-modality prior knowledge, while only the communication module is updated, so that the model can focus on learning image-layout interaction and thereby better capture the joint image-layout distribution for improved composition harmony. Our model has no design-specific inductive bias, which allows it to better preserve the original characteristics of realistic designs. We further introduce a test-time guidance strategy to enable users to impose their specific preferences on generated results. Our experiments show that, compared with prior approaches, our model can generate significantly better results in terms of image, layout and image-layout harmonization, producing outputs closer to real samples. We also demonstrate the flexibility of our model in enforcing user preferences at inference without retraining.
#03Sep 10, 2026
cs.CV
Harnessing Intrinsic Subject-Aware Attention for Controllable Multi-Subject Video Generation
Niange Yu, Ye Tian, Biaolong Chen and 5 more
Multi-subject video generation faces two key challenges: uncontrollable fidelity strength and potential semantic drift. We address these by analyzing the internal mechanisms of Diffusion Transformers (DiTs). We found that certain attention blocks naturally form an Intrinsic Spatial Grounding Map (ISGM) that precisely locates reference subjects. Building on this insight, we propose Dual-phase Intrinsic Attention Leveraging (DIAL), a framework that uses these internal signals for both training and inference. In low-noise stages, we use ISGM to guide the attention mechanism, allowing precise control over fidelity strength during inference without retraining. In high-noise stages, we use these same maps to automatically build preference pairs at no additional cost for Reinforcement Learning (RL). This RL procedure effectively anchors the model's attention to reference subjects and mitigates semantic drift. Extensive experiments show that DIAL significantly outperforms baseline models on the OpenS2V-Eval benchmark, consistently improving identity consistency and enabling controllable fidelity strength.
#04Sep 10, 2026
cs.NE
Breaking the Central Bias: Spatially Partitioned Experts for Coordinate-Based Neuroevolution
Romain Claret, Arthur Gygax, Michael O'Neill and 3 more
Evolvable-Substrate HyperNEAT (ES-HyperNEAT), a bio-inspired indirect encoding that determines neuron placement and connection weights from spatial coordinates, exhibits a failure mode on MNIST as a diagnostic benchmark. Because input pixels map to a coordinate space centered at the origin, evolved networks converge on a small central cluster of input pixels, a spatial-concentration bias; prior work observed only 21% mean accuracy in this regime. Is this bias an optimization artifact or an architectural ceiling? Inspired by Mixture-of-Experts (MoE) principles, we partition the input into non-overlapping spatial segments, each assigned to a separately evolved specialist network. With 13 such experts, this design reaches 43% mean accuracy, a 106% relative improvement over the baseline. The architectural gain does not depend on data-driven aggregation: equal-weighted averaging, which uses no validation data, already yields a 70% improvement; the gain comes from partitioning, not the weighting. Receptive-field analysis shows the mechanism: partitioning forces evolution to discover features across the entire image, expanding active pixel coverage from 4% to 79%. Absolute accuracy stays below gradient-trained baselines, but the relative gain points to central bias, not the evolutionary search. Two tools are designed to generalize beyond MNIST: a receptive-field diagnostic for silent input-coverage collapse, and a spatial-partitioning remedy that restores coverage.
#05Sep 10, 2026
cs.CV
Self-Supervised Cardiac Phase Detection via Single-Parameter Latent Orbits
John Bonnici, Matthew Baugh, Aleksandra Kulbaka and 3 more
Accurate identification of end-diastole (ED) and end-systole (ES) in echocardiography underpins the quantification of ventricular function, yet manual selection of these key frames is subjective and introduces clinically significant inter-operator variability. Recent self-supervised methods either prescribe strict periodic trajectories or learn an unconstrained low-dimensional motion subspace from reconstruction or registration objectives. The former offers interpretability but imposes restrictive assumptions on temporal progression, whereas the latter leaves cardiac phase implicit and ED/ES must be recovered through post-hoc geometric processing of the learned trajectory. We translate the physiological observation that cardiac phase is a one-dimensional signal into a prior by constraining the latent motion component to a single-parameter latent orbit, i.e., a global linear trajectory in latent space indexed by a bounded scalar phase variable. Mapping this variable through a sinusoidal nonlinearity yields an oscillatory motion signal with consistent temporal ordering, enabling direct identification of ED and ES from the learned phase signal. This inductive bias allows the model to capture an interpretable representation of the cardiac cycle, while maintaining flexibility to capture irregular heartbeats. Trained on EchoNet-Dynamic without annotations, our minimal single-parameter cardiac phase model learns an effective latent orbit, significantly improves upon the previous state of the art in ED localisation and matches it in ES localisation while using a more constrained representation and fewer training epochs. This demonstrates that a principled physiological inductive bias can match or exceed the performance of more complex representations. Code is available at: https://github.com/BonniciJ/OrbitalEcho/