graphics software developers

Abstract: Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally. Based on this insight, we introduce SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots interact locally. This allows the model to increase spectral detail without enlarging the global sequence. A spatial head further adds a local correction to the spectral prediction, while the generative state, target, and loss remain in pixel space. On ImageNet-256, SpecSlot achieves a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512, it improves FID by 2.84 over the control with 3.9% fewer FLOPs, and comes within only 0.69 FID of JiT-B/16 at 38.2% lower cost. On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.

Abstract: Unified Multimodal Models (UMMs) enable powerful multi-image synthesis but incur substantial computation from long visual contexts. While token pruning offers a direct route to accelerating inference, existing methods are primarily adapted from understanding tasks, suffering from insufficient reference image exploration, confused conditional guidance, and pooling-induced ambiguity that degrades generation quality. To address these issues, we propose CATP, the first training-free context-compression method tailored to UMM generation, which \textbf{C}alibrates the discrepancies of general \textbf{A}ttention-based \textbf{T}oken \textbf{P}runing for better quality-efficiency trade-off. Concretely, building on independent layer-wise pruning during denoising, we introduce Guidance-Consistent Pruning (GCP) to synchronize pruning decisions across conditioning branches, and a Taylor-Calibrated Attention Metric (TCAM) that derives a theoretically grounded second-order correction for pooling-induced bias. In addition, a Fast Scoring Operator (FSO) fuses the calibration into tiled online Softmax to reduce pruning cost. Extensive experiments across representative UMMs and benchmarks demonstrate that CATP consistently outperforms state-of-the-art token-reduction methods, achieving up to 2.09$\times$ attention-core speedup.

Abstract: Autoregressive (AR) image generators are powerful in principle but inflexible in practice: training with a fixed raster-scan order hardcodes a single factorization of the image distribution into the model, leaving it unable to inpaint, outpaint, or edit at inference time. Existing solutions, including the modern discrete-diffusion family, train from scratch on a mixture of token orderings, paying roughly 3x the compute of standard raster-scan training; discrete diffusion further sacrifices compatibility with the AR inference ecosystem (KV caching, vLLM) by construction. We argue that this cost is unnecessary. A pretrained raster-scan AR model has already learned the joint distribution over image tokens; what it lacks is not knowledge but the ability to express that knowledge through non-sequential conditional pathways. We propose a recipe for adapting pretrained raster-scan models into any-order generators using two components: future-aware positional embeddings that resolve a target-position identifiability problem at negligible overhead, and an ordering curriculum that traces a stable optimization path from the pretrained solution to a good any-order basin. Applied to four base models across two architectures and two scales (LlamaGen-L/XL, RAR-L/XL), the recipe substantially outperforms MaskGIT on ImageNet inpainting (FID 6.21 vs. 10.44) and outpainting, is competitive with MAR-L, trades a 1.2-1.9 increase in class-conditional FID for this flexibility, and costs 13-33% additional compute beyond pretraining, while leaving the standard AR inference stack untouched. Inpainting, outpainting, and editing, our results suggest, need not be added to autoregressive image models. They can be unlocked from the ones that exist. Code is available at https://github.com/Ordered-AR/unlockedadaptation.

Abstract: Neural representation has emerged as a powerful tool to mitigate the demand for increasingly higher-resolution images. However, a fundamental trade-off remains between structured (e.g., feature grids, which offer spatial consistency) and unstructured (e.g., 2D Gaussians, which adaptively capture local variances) representations. This choice significantly affects how well an algorithm represents an image. Here, we present Gaussian-Grid, a hybrid neural representation that adaptively integrates structured and unstructured primitives. By dynamically partitioning computation between 2D Gaussian distributions and their attached multi-resolution feature grids based on the local frequency content, Gaussian-Grid automatically optimizes its architecture to match the underlying complexity of the target signal. We evaluate our approach using high-resolution natural images and demonstrate its applications as GPU-native dataset storage, showing $1.24$-$2.14\times$ decoding speed improvement and PCIe data traffic reduction of $23\times$-$176\times$ over JPEG + ZIP data archiving. The results show that Gaussian-Grid maintains small memory footprints while achieving high reconstruction fidelity across a variety of datasets.

Abstract: Open-ended image-editing requests, such as ``decorate this room,'' admit multiple valid realizations. Direct sampling may produce ungrounded variation, whereas explicit diversity prompting often produces superficial differences. We study how to ground these requests without prematurely eliminating their legitimate ambiguity. We introduce a two-stage framework that first identifies diverse, feasible edit scopes and then realizes them as executable edit instructions. Across 300 open-ended image-editing tasks, our approach achieves 24.6% higher mean semantic dispersion and 17.1% higher mean groundedness than direct diversity prompting, while producing more valid and distinct realizations. These results show that structuring feasible edit directions before realization supports more useful grounded exploration.

Abstract: Energy-Based Models (EBMs) are a flexible class of generative models that decouple probabilistic modeling from the generation process. This decoupling directly enables compositional generation and inference-time scaling, where sample quality can be refined simply by increasing sampling steps. However, severe challenges in high-dimensional sampling and training instabilities with deep architectures have restricted EBMs to low-dimensional regimes. To resolve these bottlenecks, we introduce the Transformer Energy-Based Model (TEM), a scalable transformer energy function operating entirely within a latent space. On top of this backbone, we extract energies at multiple layers and sum them, composing complementary constraints from local texture to global semantics. Furthermore, two components make this high-quality at scale: Positive Representation Alignment (PRA), which supervises TEM's intermediate representations with a vision foundation model, and advanced sampling strategies that improve sample quality at inference. TEM-XL achieves an FID-50K of 3.82 on class-conditional ImageNet 256$\times$256 without classifier-free guidance, showing that a contrastive EBM can generate high-fidelity images. On ImageNet 512$\times$512, TEM achieves an FID-50K of 8.24, demonstrating that the approach scales to higher resolution. We further validate the core promises of EBMs: compositional generation, zero-shot inpainting, and out-of-distribution detection.

Abstract: Shadows are observable consequences of latent physical causes, arising from the interaction between illumination, object geometry, and scene structure. A model that understands shadows should be able not only to read a shadow's direction, but also to infer where a shadow should fall when it is absent, and to judge whether a shadow is physically consistent with the scene. We introduce ShadowBench, built from real photographs, which poses these three questions to vision-language models (VLMs). From each scene, we build a correct-shadow, a shadow-removed, and a human-annotated wrong-direction state. Evaluating 19 open VLMs (2B to 78B) and frontier proprietary models, we find that VLMs read a visible shadow's direction well above chance and improve with scale, yet infer the absent direction only weakly and fall to chance at judging physical consistency even under a bias-free two-alternative test that removes the yes-bias of single-image naturalness questions. Only Gemini-3.1-pro exceeds chance on validity, and even then only modestly and far below humans. Chain-of-Thought (CoT) does not help. The missing ingredient is not reasoning but knowledge that image generation models (IGMs) already hold. Routing a VLM through an IGM lifts its inference well above its direct attempt and helps it flag the wrong-direction shadows it otherwise accepts. Cast-shadow direction is a single facet of illumination understanding, but the sharp gap it reveals between VLMs and IGMs suggests image-generation or reconstruction objectives as a route toward illumination-aware VLMs.

Abstract: Training-free few-step samplers are often described as nearly quality-lossless under conventional FID/CLIP-style evaluation. It remains unclear whether that conclusion transfers to more realistic and complex generation-quality benchmarks. We first systematize the quality metrics used to evaluate training-free sampling acceleration for diffusion and flow-matching models. We find that representative training-free step-shortening methods often remain nearly lossless under FID/CLIP-class metrics, yet can suffer substantial quality degradation when evaluated with semantic-composition, dense-instruction, and long-text-rendering benchmarks such as GenEval, DPG-Bench, and LongTextBench. Parallel sampling is a notable exception: by retaining the full-step discretization and removing only the solver's temporal dependency, it can preserve near-lossless quality while accelerating sampling. Building on this observation, we propose \textit{ParaLin++}, a step-parity parallel sampler that preserves the full-step discretization while accelerating the solver iterations. Experiments across Stable Diffusion-v2, LSUN Church, FLUX.1-dev, and Ming-Flash-Omni 2.0 show that ParaLin++ preserves the full-step Euler quality profile at $2.6\times$--$4.8\times$, $1.68\times$ on FLUX, and $1.21\times$ on Ming. Project page: https://paralinplus.github.io/.

Abstract: Reproducing a published 3DGS paper takes weeks of expert engineering before any extension can be proposed. Prior autonomous-research systems read papers, draft hypotheses, and write code, but are limited by short fixed templates, low executability, and weak traceability, and none has been demonstrated end-to-end in a live computer-vision subfield with intense activity. SplatLab is an autonomous research system for 3D Gaussian Splatting that addresses these limitations. Given a research prompt, a target-paper URL, or no input at all, it reproduces a published baseline, proposes an extension, trains and ablates across Mip-NeRF 360 and LLFF, writes the manuscript, and binds every reported number to a logged training run. Five design choices distinguish it: (i) candidates mutate plugins from a verified-reproduction backbone matched to published baselines; (ii) a cascade-weighted Elo tournament across eight islands scores candidates on measured GPU PSNR with verbal-gradient feedback; (iii) a standing falsifier subtracts Elo on failure, integrating stress-test survival into selection; (iv) every reported number is bound to a logged training run, and the run cannot terminate while any claim is unbacked; (v) a seven-category 3DGS taxonomy wraps EVOLVE-BLOCK markers in domain-aligned scaffolds transferable across 3DGS repositories. On PSNR-targeted runs, SplatLab's manuscripts beat their reproduced targets by +0.18 to +3.78 dB, including +3.78 dB on astrophotographic nebula rendering, for which no prior 3DGS method exists. The research cycle compresses from months of graduate-student time to days of mostly autonomous compute. Code, all manuscripts, experiment ledgers, and the cross-run paper store are available on the project page.

Abstract: Neural Radiance Fields (NeRF) synthesize photorealistic novel views from dense image collections, but degrade sharply when few views are available. We attribute a specific, measurable defect to the objective itself: the rendering loss is a sum over independent rays and contains no term comparing densities that different rays assign to the same region of space, so nearby 3D points can settle on contradictory densities while every training pixel is reproduced correctly. We make this quantity explicit by measuring cross-ray density disagreement directly during training, and show that it stays high in vanilla NeRF, falls only incidentally under frequency-based regularization, and moves together with held-out quality in the sparse regime. We then close the gap with DriftNeRF, a drop-in regularizer that instantiates the drifting-field template on density samples: a masked exponential kernel aggregates densities from nearby points on other rays into a stop-gradient consensus target, and the resulting drifting field penalizes unsupported disagreement while leaving agreed-upon boundaries intact. The regularizer reads only sampled positions, ray identifiers, and predicted densities, so it requires no pretrained model, external geometry, or additional images, and transfers across field parameterizations. On NeRF-Synthetic it reduces residual disagreement by an order of magnitude, and across NeRF-Synthetic, DTU, and LLFF it improves 3-view PSNR over the strongest MLP-NeRF baseline by 0.42-0.65 dB and remains competitive with sparse-view Gaussian pipelines under a matched protocol without their depth, flow, or covisibility preprocessing.

Abstract: One-step distillation compresses diffusion trajectories into a single forward evaluation, yet how internal capacity should be structured within that remaining pass remains largely unexplored. In this setting, identical parameters must simultaneously preserve general generative priors and resolve prompt-specific details without iterative refinement. We formalize this challenge as functional allocation. Our analysis reveals that decoupling representations is beneficial only when expert interference outweighs shared transfer, favoring an always-active shared backbone augmented with conditionally routed residuals over standard isolated MoE architectures. We propose DeMoE, which pairs a persistent shared expert with conditionally routed expert slices across image FFNs, trained under a one-step distribution-matching objective with executed-route supervision. On one-step FLUX.2-4B generation, DeMoE achieves a GenEval score of 0.8220 while activating only 80.5% of the image-FFN width, substantially outperforming dense one-step baselines, matching the 50-step teacher overall, and surpassing it on complex composition and counting. Across COCO-10K, DeMoE establishes state-of-the-art fidelity and diversity among single-step models, demonstrating that structured internal computation reallocation provides a superior compression paradigm for diffusion models. Code is available at https://github.com/toffeecat/demoe-ttt-router.

Abstract: Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However, existing methods largely enforce a forward-only information flow: approximations produced for skipped model evaluations are immediately committed to the sampling trajectory and cannot benefit from feedback generated by later full-computation steps. We propose \textbf{PoSTCache} (\emph{Postdictive Smoothing Trajectory Cache}), a training-free and plug-and-play framework that reformulates diffusion caching as revisable trajectory inference. PoSTCache first uses an existing caching strategy to construct forward proposals for skipped steps. At a later full-computation step, the discrepancy between the refreshed prediction and its forward proposal yields a correction signal, which PoSTCache propagates backward to refine and reintegrate the preceding predictions. Viewing full-computation steps as sources of corrective feedback also changes how they should be selected. We therefore introduce the Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Experiments on representative image and video generation models show that PoSTCache consistently improves the quality--efficiency trade-off over representative caching methods.

Abstract: Visual autoregressive models produce convincing images but can misassign attributes and spatial relations. Searching multiple generation paths can improve composition, yet completing every candidate is costly. We introduce ATLAS, a training-free framework that uses the frozen generator's internal attention to guide search before final images are available. Within-prompt diagnostics identify attention-derived scores that predict final compositional preferences on Infinity, with category-dependent transfer to Switti. ATLAS uses these on-the-fly signals to progressively concentrate computation on promising trajectories. It maintains multiple coarse alternatives, prunes candidates without decoding at intermediate scales, and reserves image-space evaluation for a final choice among partial trajectories. ATLAS scales effectively on both backbones, achieving higher mean scores on the evaluated T2I-CompBench and GenEval categories as more computation is devoted to search. At the largest tested budget, the targeted T2I-CompBench mean improves by over 10% relative to each backbone. Separate profiles show over 50% less generation time than completing and evaluating all candidates at equal initial counts, alongside lower peak allocated GPU memory. Relative time savings are larger at the higher tested candidate count. Together, these results support scalable compositional search: additional sampling expands exploration, while attention-guided selection limits continuation cost without parameter updates or gradient-based optimization.

Abstract: A finite metric space with $N$ points admits an exact representation in a pseudo-Euclidean space of dimension at most $N-1$. We study the minimum dimension required for relative-error approximation of squared distances. Our main result gives an upper bound for arbitrary metrics in terms of local distance ratios: for $4/5<\varepsilon<1$, dimension $O((N/k)\log N+kρ_k^4\log N/(\varepsilon-4/5)^2)$ suffices, where $ρ_k$ is the largest ratio, over all points, between the distance to the $k$th closest point (counting the point itself) and the nearest-neighbour distance. The proof combines a low-rank landmark approximation with a sparse correction controlled by its factorization norm. For metrics with distances in $[1,Φ]$, where $Φ\ge5$ is fixed, the worst-case dimension is $Θ(N)$ for $0<\varepsilon<4/5$, lies between $Ω(\sqrt N)$ and $O(\sqrt N\log N)$ for $4/5<\varepsilon<12/13$, and is $Θ(\log N)$ for $\varepsilon>(Φ^2-1)/(Φ^2+1)$. For $Φ>5$, an intermediate range remains without matching bounds. Sign-pattern counting yields the $Ω(\sqrt N)$ lower bound for $4/5<\varepsilon<12/13$, while Hadamard matrices give explicit linear lower bounds for $\varepsilon<4/5$ via a known approximate-rank estimate. We also construct configurations with null squared intervals whose minimum representation dimension remains $N-1$ for every $\varepsilon<1$. Finally, for every fixed $0<\varepsilon<3/5$, we show that the quadratic dependence on the parameter $C$ in the dimension bound derived from Deng et al.'s pseudo-Euclidean Johnson-Lindenstrauss lemma is necessary up to a logarithmic factor.

Abstract: The human brain effortlessly deciphers colors and textures from visual stimuli and reconstructs the world around us, yet preserving such details remains challenging in image denoising. While significant progress has been made in generic image denoising, the problem of texture denoising has remained relatively under-explored. Fine details, such as intricate patterns and repetitive structures, often suffer from over-smoothing even in deep learning based approaches for denoising. In this paper, we analyze the behavior of state-of-the-art (SOTA) denoisers on texture datasets and show that performance improvements are possible only by fine-tuning the denoisers on the texture datasets. Surprisingly, fine-tuning on texture datasets result in the models catastrophically forgetting the original dataset. These findings highlight the need for specialized treatment of textures in denoising. Since denoising quality directly impacts downstream tasks such as generative modelling, we argue that dedicated texture aware denoisers are essential for faithful texture reconstruction and synthesis.

Abstract: Anatomical labels help 3D Gaussian splatting preserve vessels in cinematic renderings of CT, but the representation is still found by optimising each patient's scene, where densification decides where Gaussians go, while real-time displays cap their number. We address one part of this process: deciding where a fixed budget of B Gaussians should be placed. We propose a reusable, labelaware allocation front end that maps a patient's CT, labels and transfer function to a density over 2mm cells and draws exactly B positions from it without per-patient optimisation; the other Gaussian attributes are still fitted per patient. The front end is learned once from per-patient optimised scenes and implemented as a population-trained neural prior and an interpretable seven-parameter visibility-aware closed form (VACF). With identical attribute fitting on 20 held-out patients at 10K-100K Gaussians, both raise vessel-masked PSNR over uniform allocation by 4.5-7.1dB and over pruning a dense voxel-aligned model by 1.6-3.3dB. The neural prior does not differ significantly from per-patient optimised allocation at any budget, and VACF is within 0.6dB of it on vessels; on 19 external patients the neural prior again matches it on vessels and VACF is within 0.7dB. As an initialisation for position optimisation, VACF allocation beats per-patient densification on vessels at 10K Gaussians and, at 25K, reaches a higher vessel PSNR after 118s of optimisation than densification after 248s. Under altered rendering presets, VACF transfers to new colours and lighting but not to a label-free transfer function, whereas the neural prior shows the opposite pattern.

Abstract: We present Co-DSL, a framework in which users and LLMs co-design reusable spatial concepts for layout generation. In the first stage, our system allows users to create a library of personalized operations. To add a concept, a user provides a description, which the LLM converts into an initial implementation. The user can then observe the concept's execution behavior over example scenarios and provide lightweight feedback to the LLM, which revises the concept's implementation logic. In the second stage, given a library, an LLM can author a layout program that meets a new layout request from a user. Executing a layout program instantiates a concept-graph, a declarative composition of library concepts. In this way, we require the LLM only to specify which concepts should hold, while we use a solver to search for a layout that satisfies them. Across five spatial layout domains, we find that Co-DSL libraries let LLMs capture layout patterns that are challenging for prior fixed-operation systems, and demonstrate that our concept-graph representation is needed for LLMs to reliably compose concepts together. We further find that Co-DSL lets LLMs author generative layout programs that expose high-level parametric controls users can manipulate to explore a distribution of related layouts without re-prompting.

Abstract: 4D Gaussian Splatting (4DGS) has emerged as a powerful dynamic scene representation, supporting real-time high quality rendering but suffers from prohibitive storage overhead. Existing methods typically attempt to solve this by treating compression as a per-scene optimization problem, limiting generalizability and applicability where low-latency compression is required. In this work, we leverage existing image and video codecs and propose a generalized post-training compression framework for 4DGS to address these issues. Specifically, we apply importance-based adaptive quantization to Gaussians according to their importance levels, effectively refining bitrate allocation while remaining fully compatible with existing codecs. To further improve compression efficiency, we utilize the statistical properties of 4DGS and propose motion- and importance-aware packing to reorganize attributes into codec-friendly 2D maps. Our approach can achieve over 40x compression rate in seconds while maintaining high quality, facilitating the practical deployment of 4DGS.

Abstract: GPU working sets have outgrown GPU memory, so applications must oversubscribe it, keeping only hot data on the GPU. NVIDIA's Unified Virtual Memory (UVM) supports oversubscription, but its performance degrades once data exceeds the GPU's memory. This stems from its fault-driven data placement, wasting interconnect bandwidth, and underutilizing the GPU. This raises the question: can UVM oversubscription become practical on stock GPUs, exploiting GPU memory speed and minimizing wasted interconnect bandwidth, without placement hints, profiling runs, or application changes? To this end, we present Ember, a UVM management framework for oversubscribed workloads. Ember introduces a hybrid address space; allocations start as mapped into host memory, omitting data migration. Ember's counter-driven region promotion uses GPU's hardware counters to fetch hot regions to the GPU, its decay-based eager reclamation proactively evicts cold regions in batches, while via online region re-placement, Ember tracks the hot set at runtime. We implement Ember within NVIDIA's open-source UVM driver for unmodified applications and as a userspace variant on stock CUDA. At 1.5x oversubscription, it is 6.8-10.6x faster than UVM on six applications across three PCIe GPUs, and generates LLM tokens 4.7-16x faster at sizes where manual placement cannot run at all. Without a profiling run, it beats a profile-guided alternative by 2.1x on kernel time.

Abstract: Representation alignment with pretrained visual encoders improves diffusion transformer training but requires auxiliary encoders and extra training overhead. Recent approaches instead exploit representations from the generative model's own variational autoencoder (VAE). However, reconstruction-oriented VAE features are not optimized for the semantic discrimination and spatial structure needed for generative learning. We introduce \textbf{CoVA} (\textbf{Co}mposed \textbf{VA}E Representations), a lightweight framework that composes VAE features online into stage-wise guidance for flow-matching training. A small latent composer is jointly trained with class supervision, masked latent reconstruction, and horizontal-flip consistency, yielding semantically discriminative, spatially structured dense features and sample-dependent class posteriors. CoVA aligns intermediate SiT features with the composed representations early in training, then removes the increasingly detrimental alignment and reuses detached composer posteriors for sample-dependent soft label conditioning. We further uncover a hidden confound in soft label conditioning: averaging class embeddings attenuates their magnitude, coupling semantic interpolation with conditioning strength. This attenuation alone can substantially improve FID while degrading Recall, revealing a non-semantic source of the apparent gain. CoVA corrects this confound by restoring the hard-label magnitude, isolating posterior-dependent directional semantics. CoVA adds only 8.5M training-only auxiliary parameters (1.3\% of SiT-XL/2) and introduces no inference-time overhead. On ImageNet 256$\times$256, CoVA reduces SiT-XL/2 FID from 17.2 to 7.4 (\textbf{57\%}) in 400K steps. With classifier-free guidance, it achieves an FID of \textbf{2.03} in only 400 epochs, versus 2.06 after 1,400 epochs for vanilla SiT-XL/2.

Abstract: Visual agents for symbolic graphics reasoning depend on executable world representations, yet continued improvement remains bottlenecked by ambiguous supervision from final renders. Recent agentic systems refine scenes through repeated interaction, but they lack mechanisms to recursively extract actionable evidence from execution traces, overlooking which proposed symbolic edits survived, disappeared, or interfered during program execution. To address this, we propose VETR (Verified Execution Trajectories for Refinement), a framework that enables agentic recursive self-improvement over executable world representations for symbolic graphics reasoning. Within each task, VETR retains aligned program/render records with bounded runtime observations and introduces an execution refiner that compiles observed writes and state changes into constrained alternative programs. The refiner validates their endpoints and preservation conditions against captured runtime state; the ordinary visual verifier then assesses goal relevance. Repair construction makes no additional language-model calls, but can require additional executions. Evaluations on BlenderGym (280 tasks) and BlenderBench (27 released tasks) with GPT-5.6-luna and Qwen3.8-27B show that VETR consistently outperforms existing refinement strategies, including BlenderAlchemy and VIGA. Controlled ablations evaluate the contributions of execution refinement and role-asymmetric evidence organization under fixed generator-interaction caps.

Abstract: Most image generation models rely on uniform tokenization, allocating the exact same computational budget to equally-sized image patches. This static paradigm cannot adapt to different resource constraints at inference time, and yields suboptimal quality-cost tradeoff by devoting the same effort to both plain backgrounds and intricate details. We propose BudgetPix, an adaptive tokenization framework that dynamically allocates compute based on visual complexity and spatial layout, enabling flexible computational budgeting at inference time. BudgetPix comprises three key components: (1) an adaptive encoder that maps a fixed-size image to a variable-length token sequence using an entropy-guided quadtree alongside a multi-scale patch embedder; (2) a scale-aware decoder reconstructs fixed-resolution images from multi-scale token sets; and (3) a flexible training and sampling schedule that enables pixel-space denoisers to operate across variable token counts. BudgetPix seamlessly integrates with existing pixel-space diffusion architectures, enabling a single checkpoint to be operated at a wide range of compute budgets. Evaluated on text-to-image generation, BudgetPix matches the fidelity of MiniT2I-L at $512^2$ and PixelDiT at $1024^2$ using just 25% of the original compute budget. In class-conditional generation using a MeanFlow backbone, BudgetPix requires merely 60% of the full compute budget to produce images with near-zero quality degradation, observing a marginal 0.8-point increase in FID. Comprehensive assessments by human and VLM judges confirm that BudgetPix establishes a significantly improved quality-efficiency tradeoff over prior budget-adaptive baselines. More details are available at our project page: https://karaozgur.com/BudgetPix

Abstract: Co-denoising pretrained representations such as DINO can substantially improve the training speed and quality of flow matching models, but it introduces a second denoising trajectory and requires carefully designed schedules. We propose a simpler alternative: predict the pretrained representation directly, then condition the model on its own prediction. This removes the need for a second ODE and any representation-specific denoising schedules, while retaining the benefits of representation guidance. Our approach converges substantially faster and achieves better generation quality as measured by FID score. On ImageNet, it outperforms the state of the art in latent space at 2x fewer epochs than prior methods; in pixel space, it improves FID over comparable prior methods by more than 20%. These results support a simple principle: do not denoise what you can predict. Our code is openly available at https://github.com/arijit-hub/dino_forcing.

Abstract: This paper studies the problem of learning disentangled representations of objects and their attributes from raw, unstructured image data. Slot-based methods have shown considerable success in unsupervised learning of object representations from images. Block-slot attention-based methods extend this framework to attribute representations by assuming a uniform factorization of object representations into attributes, which may be suboptimal and consequently limit the quality of the learned representations. We therefore investigate a framework for jointly discovering object and attribute representations. Our key contribution is leveraging the Linear Representation Hypothesis (LRH), which postulates that composable concepts can be represented as linearly additive subspaces in slot representations. Based on this insight, we propose a probabilistic model connecting images, slots (objects), and blocks (attributes). We present an architecture that leverages block attention to connect attribute representations to slots and incorporates LRH in both object and attribute representation spaces. This architecture effectively optimizes the Evidence Lower Bound (ELBO) of the proposed graphical model. Our experiments demonstrate (i) effective discovery of disentangled object and attribute representations, (ii) empirical evidence for LRH in slot space, and (iii) the ability to perform image editing owing to the disentangled and interpretable nature of the learned representations. Our experiments on multiple datasets demonstrate improvements in DCI scores over state-of-the-art methods.

Abstract: Diffusion Transformers (DiTs) achieve remarkable performance in video synthesis, but their iterative denoising process suffers from high inference latency. To address this, caching has emerged as an effective acceleration strategy by capitalizing on inter-step redundancy during denoising. Existing dynamic caching methods typically estimate the error that cache reuse would introduce at each denoising step (step error) to guide cache decisions, whereas our concern is how much quality loss cache reuse would cause in the final generated video (terminal error). We show that step error does not directly correspond to terminal error and that latent information helps capture their relationship, thereby informing cache decisions. Moreover, existing threshold-based methods cannot provide precise speedup control, making it difficult to meet practical requirements for user-specified acceleration targets. To address these limitations, we introduce MORCA, a cache scheduling framework trained through offline-to-online reinforcement learning to make latent-aware reuse/recompute decisions under user-specified acceleration targets. Extensive experiments on different video generation models across multiple target acceleration ratios demonstrate that MORCA achieves better generation fidelity than state-of-the-art caching methods under comparable computational budgets. Code is available at https://github.com/x10ngyx/MORCA.

Abstract: Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256$\times$256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.

Abstract: Recent diffusion-based image generation backbones have grown substantially in scale, making the network inference cost increase rapidly. While diffusion distillation techniques can reduce the number of inference steps, high-quality image generation within a single full-backbone-forward compute budget remains challenging. Existing one-step methods typically allocate this budget to a single evaluation of a monolithic student. However, approximating the heterogeneous coarse-to-fine transport with a single monolithic mapping is difficult and often leads to over-smoothed outputs. To address this issue, we propose Phase-wise Velocity Distillation (PVD), which partitions the generation timeline into a coarse and a fine phase, and models the transition within each phase via the average velocity. A dedicated half-sized expert is assigned to each phase, decoupling structural composition from detail refinement while keeping the cumulative computation equivalent to one full-backbone forward pass. We show that the use of two half-sized phase-specific experts outperforms a single full-size monolithic student. On class-conditional image generation, PVD achieves an FID of 1.48 on ImageNet 256 x 256. On more complex text-to-image (T2I) tasks, PVD-distilled models (Stable Diffusion 3.5-Medium, FLUX.1-dev, Qwen-Image) produce results competitive with their multi-step teachers, significantly outperforming prior distillation methods. Moreover, across the evaluated T2I backbones, PVD reduces active parameters by 49.10-50.89% and peak VRAM by 45.76-48.36% compared to the corresponding teachers. Source code and distilled models are available at https://github.com/PolyU-VCLab/PVD.

Abstract: Image and video diffusion models allocate equal computation to every region, even when the intended scene calls for varying levels of detail. The spatial distribution of detail can often be anticipated before generation, indicating where computation can be reduced. We introduce Level-of-Token (LoT) Diffusion, a framework that turns this knowledge into an explicit multiresolution token layout (Level-of-Token layout) for adaptive and efficient generation. Tokens represent rectangular patches of varying sizes and shapes, allocating finer tokens where detail is needed and coarser tokens elsewhere. We adapt pretrained diffusion transformers to LoT layouts through a patch-wise asymmetric flow parametrization and embeddings for multiresolution tokens, preserving full-resolution flow prediction at every denoising step while processing only a reduced token sequence. LoT Diffusion enables layout-adaptive generation while preserving pretrained generative priors. We demonstrate LoT with layouts derived from semantic masks, bounding boxes, texture variance, and depth-of-field cues, as well as agentic plans. Across image and video generation, LoT offers favorable quality-efficiency tradeoffs, with significant speedups determined by the layout's token budget. Our project website is at https://georgenakayama.github.io/lotdiffusion/.

Abstract: Strong image generation models are conditioned on class labels, aligned to frozen pretrained encoders, or built on separately trained autoencoders. While effective, generation then depends on supervision or pretraining: labels must be annotated, and encoders or autoencoders pretrained for the target domain. We study joint generative and self-supervised representation learning in a single model, enabling self-conditioned generation without labels or pretrained models. This is challenging because the objectives are mismatched: contrastive learning consumes clean augmented views and favors coarse, invariant semantics, while flow matching consumes noisy images and must preserve the fine detail and spatial layout that contrastive learning discards. We propose SCION (Self-conditioned Generation on Self-supervised representation), whose core is a single pixel-space encoder conditioned on the flow timestep and an embedding. For representation learning, this conditioning embedding is a learned global vector shared across images, with the encoder's [CLS] token yielding the semantic representation trained by the contrastive loss. For generative training, the conditioning embedding is the image's own [CLS] representation, while patch tokens pass through a decoder to predict the image. To sample without a reference image at inference, we jointly learn a prior over the embedding. Gradient-norm balancing and stop-gradient mechanisms enable joint optimization in one run. SCION is self-supervised and self-contained, with no labels or pretrained models. On ImageNet 256x256, with the JiT-B recipe and no representation guidance, SCION reaches 8.92 FID, surpassing class-unconditional iREPA, which aligns to pretrained DINOv2 (46.44), and RCG, which conditions on it (14.27). With JiT-L, SCION achieves 5.89 FID without guidance and 3.47 with representation guidance, outperforming RCG with the ADM recipe (6.24).

Abstract: Stochastic order-independent transparency enables efficient and elegant rendering of primitive-based radiance fields like 3D Gaussian Splatting models, but remains impractical due to the inherent visible noise in the output. We propose a principled approach to minimize high-frequency noise, addressing its sources at the representation and image synthesis level. During stochastic rendering, our history-based spatial resampling scheme drastically accelerates image convergence, while temporal importance resampling ensures coherence under camera movement. During training, a color regularizer implicitly reduces the variance along view rays in the 3DGS models. With these properties, our optimized, Vulkan-based renderer effectively mitigates output noise at low and high sample counts, achieving a substantial 13~dB PSNR increase in quality over previous stochastic methods at 1 sample per pixel and quickly converging to sorted 3DGS with an average L1 error of less than $10^{-4}$.

Abstract: Fine-grained image editing requires more than producing a visually plausible result: an editor must execute the requested attribute change precisely while leaving everything else intact. However, existing benchmarks leave a critical gap between realism and verifiability: benchmarks built on realistic images typically rely on human or vision--language model judgments, while deterministic evaluation has largely focused on synthetic shape canvases, with application-oriented extensions primarily limited to charts. This makes it difficult to determine precisely how much of a requested edit was executed, where unintended changes occurred, and whether small differences between models reflect genuine editing capability or evaluator uncertainty. To bridge this gap, we present VeriEdit-Bench, a benchmark for fine-grained, instruction-faithful image editing across realistic structured assets with deterministic, four-axis evaluation. Its 1,740 cases are compiled from the source code of 153 Scalable Vector Graphics (SVG) graphics, charts, web interfaces, and presentation slides. Controlled source-code edits preserve the original visual context while yielding exact target images, pixel-level edit masks, and explicit edit specifications, enabling reproducible scoring along four axes: edit fidelity, preservation, localization, and magnitude. Evaluating eleven editors, we find that even the strongest model remains far from full credit; rankings for the same recoloring operation reverse between charts and SVG graphics; and outputs with similar pixel-accuracy profiles can still differ substantially in localization and change magnitude. This decomposition yields graded, verifiable feedback and exposes model-specific capability and failure profiles that holistic scores or evaluator-dependent judgments may obscure.

Abstract: Diffusion Transformers (DiTs) can generate high-quality images and videos, but generating each sample requires multiple costly DiT forward passes. Two common ways to accelerate DiT sampling are step distillation, which reduces the number of sampling steps, and caching, which skips some DiT evaluations by reusing a tensor computed at an earlier step. Most caching methods decide in advance which tensor to reuse. After distillation, adjacent sampling steps are farther apart. Reusing a tensor across this larger gap introduces more error, so choosing what to cache becomes especially important. We therefore introduce AutoTarget, a method that chooses the cached tensor for a given model, solver, and reuse schedule. AutoTarget uses a small set of runs without cache reuse to measure the error caused by reusing each candidate tensor, then selects the candidate with the lowest error. We also analyze how an error at one reuse step affects the final sample. For Euler sampling, we identify cache targets that produce the same trajectory and show why a stored solver update may not. Experiments on distilled image and video DiTs show that the best cache target changes with the model, image resolution, and solver. AutoTarget reduces DiT evaluations and retained cache storage. Generation quality remains close to the corresponding uncached run. On the tested PixArt-LCM and FLUX.1-schnell settings, its calibration ranking matches the ranking from held-out cached runs. To help others reproduce the method, we provide its core implementation on GitHub at https://github.com/wali1024-offical/AutoTarget.

Abstract: We present Windfoil, a GPU-friendly algorithm that treats rasterisation and differentiable vector graphics as two sides of the same problem by evaluating the box-filtered winding number of quadratic Bézier contours in closed form. We implement this in WebGPU, allowing it to run across a range of environments, including a web browser on a consumer laptop, and apply the system to real-time 2D rendering, high-resolution rasterisation for print media, and a differentiable renderer. We compare our renderer against Skia, a production-grade engine, and Slug, a popular GPU rasterisation algorithm for games and real-time applications, measuring fidelity to a reference box-filtered coverage. Our renderer matches the reference more closely than either, at performance comparable to Slug. We also compare our optimiser against DiffVG and Bézier Splatting, where it reaches equivalent or better reconstruction quality at a fraction of the per-step cost, scaling to tens of thousands of shapes at interactive rates.

Abstract: Latent Diffusion Models (LDMs) typically adopt a two-stage pipeline: an auto-encoder (AE) is first pre-trained to define a latent space, then a diffusion model is trained to perform denoising within it. Such a two-stage design introduces a representation mismatch, as the latent space is optimized for reconstruction rather than adapting the denoising dynamics. We reveal that the LDM itself is an AE, and consequently present LDM-is-AE, an end-to-end one-stage LDM training framework that eliminates the need for a separately trained tokenizer. Our key observation is that the LDM backbone actually performs a latent-to-feature-to-latent transformation at each denoising step, which can be interpreted as an internal decoding--encoding process. Leveraging this structure, we split the DiT backbone into two reciprocal components, DiT-E (i.e., DiT Encoding) and DiT-D (i.e., DiT Decoding), and impose image-space supervision on the intermediate features across all timesteps. Our model encourages the internal representation to align with the image domain throughout denoising, thereby establishing an explicit latent-to-image-to-latent path. At the zero-noise timestep, our model further performs an image-to-latent-to-image mapping, corresponding to an auto-encoding process. As a result, LDM-is-AE jointly learns latent representations and denoising dynamics in an end-to-end manner, yielding a diffusion-native latent space tailored to the generation process. Experiments demonstrate that LDM-is-AE exhibits highly competitive generation performance, achieving an FID of 1.80 and 1.90 on 256x256 and 512x512 class-conditional image generation, respectively.

Abstract: Weight compression helps large neural networks fit deployment memory budgets, but common fixed-width formats offer only coarse storage choices. Entropy coding supports finer rates, yet the achieved size depends on the quantized weight distribution and coding overhead. Exploiting this flexibility requires accurate rate selection and efficient weight reconstruction for inference. We present EntroPack, an entropy-coded weight compressor that supports arbitrary target bitrates without activation calibration or fine-tuning. It combines row-normalized $E_8$ lattice quantization with a conditional probability model of lattice coordinates. Sampled storage estimates select the quantization resolution without repeated full-stream encoding. The final coordinates are entropy-coded in independently decodable tiles, enabling fast, fused symbol decoding and numerical weight reconstruction on the GPU. EntroPack supports floating-point and integer weight containers, such as BF16, FP16, FP8, and INT8, with storage bitrate controlled independently of numerical precision. Online decoding adds latency that grows with weight count, making the method well suited to compute-intensive workloads such as diffusion denoising and Transformer prefill. Experiments demonstrate fast encoding and modest inference overhead in these settings. When compressing the linear-layer weights of the image generator Z-Image-Turbo, EntroPack achieves substantially lower weight and denoiser output errors than fixed-width formats at comparable storage rates, with modest denoising-step overhead. Targeting 4 bits per parameter, it achieves lower weight and denoiser output errors than NF4, including about 24% lower relative $L_2$ weight error, with less storage. Source code is available at https://github.com/modelscope/entropack.

Abstract: High-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's current capabilities, and the inherent trade-off between kernel correctness and performance. To address these challenges, we propose KernelZero, a co-evolution framework that continuously improves GPU kernel generation through two specialized models: a Proposer that generates Torch modules from API sets and a Coder that translates them into CUDA or Triton kernels. KernelZero uses a frontier-driven module generation mechanism to continuously produce capability-aligned training modules based on the Coder's current weaknesses. It further introduces Correctness-Aware Group Relative Policy Optimization (CA-GRPO), which optimizes performance only after correctness becomes sufficiently reliable. By alternating the optimization of the Proposer and Coder, KernelZero forms an automatic curriculum that enables targeted and training-efficient capability improvement. Empirically, KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton. On KernelBench Level 1 and 2, it achieves CUDA pass@1 scores of 75.8% and 69.6%, respectively, with pass@10 reaching 100% and 97%. On Triton, it achieves pass@1 scores of 77.2% and 72.5%, respectively.

Abstract: Iterative generation poses a joint optimization problem across steps, as intermediate predictions shape subsequent computations and ultimately determine the final output distribution. Few-step generators distilled from pretrained diffusion and flow-matching models make such optimization computationally practical end to end. We build on this opportunity with a distill-then-refine approach that uses teacher imitation to establish a strong initialization for a few-step rollout in latent space, then shifts to end-to-end refinement of the complete latent rollout against real data. We introduce FiST (Flow-in-Stage Transformer), an architecture that composes learned latent-state transitions in a few stages using a shared Transformer, with optional cross-stage hidden communication. Distillation applies teacher-forced regression to selected states along teacher trajectories; refinement replaces this supervision with adversarial and auxiliary classification objectives on the final latent output. A trainable discriminator module operates on semantically rich features extracted from clean real and generated latents by a frozen SiT backbone pretrained with REPA. All training takes place in latent space, without image decoding. During refinement, FiST consumes its own intermediate predictions, and endpoint gradients pass through every generation stage. For class-conditional generation on ImageNet at $256\times256$, our approach achieves FID 1.11 (IS 282) with three stages and FID 1.15 (IS 280) with two. These results demonstrate competitive few-step generation through learned distribution-level supervision, without explicit Fréchet-distance minimization. Ablations characterize how distillation, pretrained checkpoint choices, refinement supervision, and cross-stage hidden communication affect generation quality.

Abstract: How phenotypic transformations are implemented by changes in underlying regulatory dynamics remains a central question in developmental biology. Inspired by D'Arcy Thompson's 1917 "On Growth and Form", we ask whether coherent large-scale transformations of morphology can be encoded as low-dimensional modulations of a self-organizing developmental system. We use neural cellular automata (NCAs) as bio-inspired models of distributed development, in which a shared local regulatory network grows target morphologies from a single cell. We apply low-rank adaptation (LoRA) to pretrained NCAs, representing each adapted developmental program as a low-rank modulation of a fixed regulatory scaffold. Horizontal and vertical scaling of a fully grown 2D emoji phenotype can each be implemented by rank-one adaptations. Their linear combinations parametrically control phenotype size, generalize beyond the training distribution, and compose with target-specific adapters. Strikingly, adaptations learned for one phenotype transfer zero-shot across structurally and semantically diverse phenotypes sharing the same reference scaffold, while largely preserving internal features. This suggests reusable system-level hyper-directions of scale rather than morphology-specific transformations. From approximately 25,000 independently trained phenotype-specific NCA adapters with a shared scaffold, we further identify latent low-dimensional directions that functionally control phenotypic variation including scaling, style, and symmetrical fission. Together, our results provide a computational realization of D'Arcy Thompson's remarkable grid transformations in a 2D NCA---a minimal cybernetic tissue in which variations of fully grown emoji phenotypes can be encoded, combined, and controlled through low-dimensional directions in regulatory weight space.

Abstract: Most of the memory of a 3D Gaussian Splatting model holds spherical-harmonic colour coefficients, yet each Gaussian is seen only from the narrow cone of directions of the training cameras. We turn this into a distortion metric that other compressors can adopt: a per-Gaussian observation Gram matrix, accumulated from viewing directions and blending weights, is the exact first-order map from coefficient changes to squared image error and needs only the model and the camera poses. Under it, degree reduction becomes a closed-form projection that generalises truncation, degree allocation a Lagrangian rate-distortion problem, and vector quantisation the matrix-weighted Lloyd algorithm, of which Compressed3D's quantiser is the scalar case. Swapped into Compressed3D with everything else unchanged, the metric raises PSNR by +0.49 dB before fine-tuning, with SSIM and LPIPS following, and at matched rate still gains +0.32 dB without a single training image. A training-free stack built on the metric alone is 15% smaller than the image-free GSICO at equal quality on Mip-NeRF 360.

Abstract: Few-step autoregressive (AR) video diffusion enables low-latency streaming generation, but existing post-training methods predominantly rely on Distribution Matching Distillation (DMD), requiring both a large pretrained teacher and an online critic to estimate distributional discrepancies through diffusion scores. In this work, we ask whether this resource-intensive teacher--critic stack can be eliminated by post-training only the generator against a precomputed target distribution. Drawing inspiration from representation distribution matching (RDM) for one-step image generation, we systematically study its transfer to few-step causal video generation and identify three key barriers: a memory-intractable gradient path, a distinct video optimization regime, and representation distributions that underconstrain temporal dynamics. We introduce ViRDM, a teacher- and critic-free video post-training recipe that addresses these barriers sequentially. By coupling RDM with stochastically truncated clean-exit supervision, a lightweight VAE decoder, and staged vector--Jacobian products, ViRDM makes representation distribution matching memory-feasible for multi-step causal video rollouts. We further establish effective generated-population and initialization regimes for video RDM, and introduce lightweight dynamics regularization to compensate for the underconstrained temporal dynamics. ViRDM turns three-network distillation into generator-only post-training, reducing GPU memory use and training time while improving video quality. With only 20 generator updates, the recipe reaches 84.87 on the official VBench evaluation, outperforming the previous best few-step causal baseline by 0.36, while requiring 16 A100 GPU-hours. We additionally report exploratory results demonstrating the potential of the same recipe for lower causal sampling budget and for one-, two-, and four-step bidirectional generation.

Abstract: Recent advances in pixel-space diffusion models have narrowed the image quality gap with latent-space diffusion, but still converge more slowly and lag behind in final image quality. We argue that a key reason is the lack of an explicit representation prior: unlike latent diffusion, which usually denoises in a compact and structured latent space, pixel diffusion needs to learn denoising-friendly representations and pixel generation simultaneously from raw RGB space. To address this problem, we propose PixelDiT2, an end-to-end pixel-space diffusion model designed to decouple representation learning from pixel generation without introducing an autoencoder or latent reconstruction bottleneck. We propose representation grounding that uses a frozen pretrained vision foundation model to provide explicit per-patch representation guidance throughout denoising, allowing the pixel diffusion transformer to focus more on pixel generation. On ImageNet-256x256, PixelDiT2 achieves an FID of 1.46 after 600 epochs; at 512x512 resolution, PixelDiT2 achieves an FID of 1.48 after 680 epochs.

Abstract: Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost, and a deployable low-bit kernel must be accurate and fast. Accuracy is limited by outliers: a block's quantization scale is set by its largest entries, leaving typical entries confined to a narrow range of representable values. Prior work smooths queries and keys, but value outliers follow no fixed channel or spatiotemporal structure and remain the dominant source of output error. Speed is limited by softmax: low-bit Tensor Cores accelerate only the two matrix multiplications, so the high-precision exponential between them becomes the longest pipeline stage on datacenter GPUs. We propose VC-Attention, a training-free low-bit attention framework that addresses both by pairing Value smoothing with a fused probability Cast. V-Smooth reorders value tokens by lightweight online clustering, so the tokens in a hardware block quantize well together. It quantizes only the residual after subtracting the block mean, and restores that mean from the row sum the online softmax already maintains. ExpCast-FP8 maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion. We implement VC-Attention for B200, B300, H200, RTX PRO 6000, and RTX 5090. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5, and MiniMax-H3, VC-Attention improves fidelity over low-bit baselines, speeds up the attention kernel over BF16 FlashAttention-4 by 1.46-1.59x on datacenter Blackwell and Hopper and by 2.3-3.6x on workstation cards, and generates a clip 1.13-1.19x and 1.36-1.70x faster end to end.

Abstract: Few-step distillation accelerates diffusion models but must balance diversity and fidelity: trajectory-based distillation preserves mode coverage, while distribution matching sharpens samples but can reduce diversity. We show that this tension can be exploited in a noise-regime-dependent way: high-noise steps largely determine global modes, whereas low-noise steps refine local details. We propose CrossDistill, a trajectory-level hybrid distillation framework that splits the sampling trajectory at a crossover point, applies a trajectory-preserving objective on the high-noise interval and a distribution-matching objective on the low-noise interval, and couples the two stages through the crossover state. In contrast to loss-level mixing, and complementarily to training-time two-stage recipes, CrossDistill explicitly assigns complementary objectives along the noise axis, so that global branching is preserved before local statistics are sharpened. CrossDistill is a noise-level scheduling policy: PCM and DMD are plug-in instantiations, while the noise partition, crossover coupling, and objective ordering are the key design elements. Experiments on text-to-video diffusion models and qualitative image-to-video results show that CrossDistill expands the few-step quality-diversity frontier, retaining seed-level variation while achieving competitive visual fidelity.

Abstract: Video tokenizers have emerged as a cornerstone of modern video modeling, underpinning progress in compression, reconstruction and generation by mapping high-dimensional visual signals into compact latent spaces. However, despite this progress, current tokenization paradigms largely remain within the 2D visual domain, treating videos as image sequences rather than observations of an underlying dynamic 3D world. Consequently, the learned tokens inherit this observation-centric bias, limiting their capacity to compactly represent real-world 4D scenes. To mitigate this issue, we propose VideoTok4D, a novel 4D-aware video tokenizer for compact world representation. Specifically, our approach comprises three key designs: 1) a spatiotemporal disentanglement strategy that factorizes videos into static and dynamic tokens for holistic world modeling; 2) a track-aware dynamic attention mechanism that aggregates trajectory-aligned cues to promote cross-view motion consistency; and 3) Co4DGen, a diffusion prior learned over the resulting VideoTok4D token space for efficient 4D scene generation. Extensive experiments have demonstrated that our proposed method achieves state-of-the-art performance while requiring up to 4 orders of magnitude less storage than dense 4D representations. Moreover, the compact token space substantially shortens diffusion sequences, enabling efficient generation.

Abstract: Diffusion and flow-matching schedules control the signal and noise coefficients that mix data and noise along affine probability paths. Minimizing a kinetic action defined on coefficient paths, motivated by optimal transport, helps explain strong baselines but remains model-agnostic and ignores prediction error. Here we introduce a model-aware schedule construction based on fiberwise optimal transport. At a fixed time and state on the probability path, compatible signal/noise decompositions form an affine fiber. We define a fiberwise prediction risk by averaging optimal-transport costs between the true and predictor-induced decompositions within these fibers. On a fixed coefficient curve, combining this risk with coefficient-path kinetic action yields a closed-form optimal time allocation. This construction extends to general linear prediction targets, and the risk profile can be estimated from an early baseline checkpoint. We evaluate DDPMs and flow matching across prediction targets, training configurations, risk-estimation checkpoints, datasets, and architectures. Our model-aware schedules consistently outperform strong baselines, including a 38.6% relative FID reduction for flow matching on CIFAR-10 at 16 function evaluations. Each model-agnostic kinetic baseline determines its own kinetic reference coordinate. In these coordinates, fiberwise-risk profiles from independently trained models in different settings align closely after normalization to unit area. The resulting schedule deformations used in training also align, suggesting empirical universality across the evaluated models and settings. Pretrained-checkpoint diagnostics extend this normalized-risk agreement to larger conditional latent diffusion and 2-RF models. A frozen analytic allocation template retains most of the model-aware improvement without further risk estimation or model-specific fitting.