arXiv subjects

Zihao Fan

Publications and source records attributed to Zihao Fan.

17 recordsLinked to original sources

arXiv2026

OmniOPD: Stable On-Policy Distillation for Joint Audio-Video Generation

Reinforcement learning (RL) elevates joint audio-video generation, but optimizing heterogeneous, modality-specific audio and video rewards often introduces cross-modal interference. On-policy distillation (OPD) alleviates this issue by transferring the capabilities of specialized audio and video RL teachers into a single student model. However, a significant imbalance in cross-modal token counts leaves audio with sparse supervision, causing unstable distillation. We observe that naively increasing the audio loss weight fails to solve this problem since audio learning relies heavily on visual context, indicating richer visual cues are crucial to offset sparse audio signals. To address this, we propose OmniOPD, a modality-wise on-policy distillation framework designed for balanced and stable audio-video generation. First, we construct privileged visual information using asymmetric noise scheduling, allowing the audio teacher to generate more informative targets from cleaner video context. Second, we introduce grouped multi-noise sampling to expand audio supervision coverage and stabilize training. Third, we introduce timestep-wise importance weighting based on teacher-student transition divergence, preserving reliable supervision while down-weighting unreliable learning signals, to stabilize fluctuations in audio learning. Evaluated on JavisBench and VBench demonstrate OmniOPD preserves audio quality while significantly enhancing video generation, achieving effective joint distillation.

cs.CV↗

arXiv2026

Scaling LLM Agent Tool Access: The MCP Gateway

LLM agents increasingly rely on tool calling to complete tasks. Model Context Protocol (MCP) standardizes tool calling between LLM agents and tool providers. It reduces integration cost and drives rapid growth in tool scale. In production deployments, tool access becomes difficult to scale. On the tool provider side, legacy services are not directly callable through MCP; the rapid protocol development creates ongoing compatibility cost. On the agent side, the number of mounted tools is limited by the LLM context window and inference overhead; mounting a large tool set increases token usage and inference latency and can reduce task success rate. Moreover, for stateful MCP backends with multiple replicas, preserving session affinity increases client-side complexity. We present a cloud-scale MCP Gateway system. It breaks the direct-connect model on the data plane and offloads legacy service integration, consolidating incompatible MCP variants, access control, tool recommendation, and session-aware routing at the gateway. The conversion module supports 315 OpenAPI services in the cloud and integrates MCP transport variants. Session-aware routing decouples the scalability of the gateway layer and the MCP service layer. The hybrid retrieval engine sustains $\sim$98\% Top-15 recall; it scales agent tool access to 3{,}000+ tools with improved tool selection accuracy. We also share the lessons learned from deploying the gateway system in production.

cs.DC↗

arXiv2026

Direct Forcing: Single-Stage Training with Representation-Space Supervision for Few-Step Autoregressive Video Generation

Few-step autoregressive video generation enables low-latency streaming, yet prevailing training pipelines separate causal adaptation from distribution matching. Recent representation-based text-to-video methods retain this separation by starting with an ODE initialization stage, even when the target distribution is defined directly by reference videos. We introduce \method, the first single-stage framework that uses reference videos both as paired examples of how a history continues and as a target population for self-generated rollouts. Specifically, Direct Forcing is trained with two complementary signals: (1) Causal Continuation Alignment learns paired spatiotemporal correspondence under reference histories; and (2) Rollout Distribution Matching employs maximum mean discrepancy to align the representation distributions of self-generated and reference videos. To coordinate these signals, we introduce Dynamic Gradient Balancing, which adapts their relative gradient contributions during training. Experiments with Wan2.1-1.3B show that Direct Forcing learns a four-step autoregressive generator in 250 updates without a separate ODE initialization stage or online diffusion-score supervision, achieving 84.63 VBench Total and 86.01 Quality.

cs.CV↗

arXiv2026

Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions

While recent autoregressive video diffusion models achieve remarkable streaming quality, they remain confined to low resolutions (e.g., 480P), leaving efficient, scalable, real-time high-resolution video generation a fundamental open challenge. To bridge this gap, we present Ultra Flash, a cascaded streaming framework capable of real-time high-resolution video generation. Ultra Flash achieves ~30 FPS at 1K resolution and ~18 FPS at 2K resolution on a single GPU through three key contributions: (1) an architecture-preserving T2V-to-TV2V super-resolution training paradigm coupled with an AIGC-oriented data degradation pipeline that effectively preserves the generative capability of the base model, enabling enhanced high-resolution detail when cascaded after mainstream low-resolution generative models; (2) a causal streaming latent upsampler paired with a high-resolution decoder, which enhances spatiotemporal coherence while enabling efficient latent spatial scaling and precise high-resolution decoding with negligible computational overhead; and (3) a cascade high-resolution streaming video generation optimization scheme that first performs hybrid-reward-enhanced sparse causalization and single-step distillation of the super-resolution model, then introduces cascaded streaming self-forcing preference optimization with dynamic cache management, jointly enhancing overall coherence, improving quality, and enabling real-time high-resolution streaming video generation. Extensive experiments demonstrate that Ultra Flash reliably produces ultra-high-resolution streaming video while maintaining state-of-the-art visual quality and superior efficiency. Project Page: https://xin1u.github.io/UltraFlash/

cs.CV↗

arXiv2026

What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling

World action models (WAMs) predict the future alongside actions during training. Due to the heavy computation cost of video denoising, whether the future must still be generated during inference is disputed: Explicit WAMs denoise it into clean frames along with every action chunk, whereas Latent WAMs discard it entirely for acceleration. We find that latent WAMs, despite matching explicit ones on in-distribution tasks, fail to retain the generalization benefits that originally motivated WAMs. To demonstrate this, we evaluate generalization along three axes: environmental perturbation, data efficiency, and task generalization. Controlled comparisons with a matched backbone, training data, and budget reveal consistent degradation across all three axes when the action expert no longer conditions on future representations. Further analysis shows that the gap arises almost entirely from the first denoising step: the benefit comes from preparing the future, not generating it. We therefore propose Simple-WAM, which simplifies future modeling into a single forward pass of fully noised video tokens and adapts the training-time noise schedule to this inference behavior. Across simulation and real-world tasks, Simple-WAM achieves the best of both worlds, leading explicit WAMs in generalization performance with efficiency comparable to Latent WAMs. Project Page: https://zrporz.github.io/Simple-WAM-Web/

cs.CV↗

arXiv2026

OmniVR: Audio-Video Conditional Generation for Archival Footage Restoration

Archival footage often suffers from coupled visual and acoustic degradations, yet most restoration systems process the two modalities separately. To address this problem, we present OmniVR, the first systematic framework for joint audio-video restoration, covering data construction, model adaptation, efficient inference, and evaluation. We construct a high-quality audio-video corpus with detailed captions and use a joint degradation pipeline to produce aligned clean and degraded pairs. Using these pairs, we adapt a pretrained text-to-audio-video model (T2AV) by introducing degraded audio-video conditions (TAV2AV), then progressively replace sample captions with a fixed restoration prompt while retaining caption/null rehearsal. The resulting AV2AV model requires no user-provided text. Under a compatible residual-learning model, we prove that this condition-annealing schedule reduces gradient variance and expected restoration risk relative to direct fixed-prompt adaptation at the same training budget. For efficient deployment, OmniVR-Flash combines reduced-resolution video conditioning, MeanFlow-based one-step distillation, and Turbo VAE, achieving approximately 38 fps at 1K and 18 fps at 2K on a single B200 GPU. We further introduce OmniVRBench to evaluate four complementary dimensions: visual quality, audio quality, temporal consistency, and audio-visual synchrony. OmniVR achieves state-of-the-art results on public benchmarks and OmniVRBench. Data, code, and model weights will be released. Project Page: https://xin1u.github.io/OminiVR_PAGE/

cs.CV↗

arXiv2026

HyDra: Demystifying and Taming Dynamic Context Parallelism at Production Scale

Long-context training runs on sequences whose lengths span orders of magnitude, and dynamic context parallelism (DCP) gives each sequence its own CP degree. Existing DCP systems either do not scale or perform poorly on mainstream models, leaving Megatron-Core (Mcore) DCP as the only option at production scale. Mcore DCP, however, sizes each degree to fit memory, which grows linearly with length while attention grows quadratically, so comparable token counts hide unequal computation. In our production 256K-context training job on more than 11K GPUs under Mcore DCP, per-rank microbatch times differ by up to 5x at similar token counts. The skew leads to a 46% pipeline bubble and a 13% data-parallel bubble. We present HyDra, a scalable load-driven DCP system. Its scheduler balances computation by pulling every rank toward one load target, and balance in turn makes that target solvable in closed form. It places sequences with lazy heaps rather than whole-pool scans. That balance asks for CP degrees larger than memory requires, so its CP engine nests an inner Ulysses group in a shallow outer ring, letting a higher degree lower computation and communication together. Evaluation at both scales shows consistent gains. On a 512-GPU testbed, HyDra raises throughput over Mcore DCP by 1.18x on average at 32K context and 2.48x at 256K. On a 2,048-GPU production job, it shrinks the pipeline bubble from 36% to 14%, cuts scheduling time by 2.3x, and raises throughput by 1.10-1.43x (avg. 1.25x) over Mcore DCP and 1.33-1.90x (avg. 1.59x) over static CP.

cs.DC↗

arXiv2026

DyMD: Preserving Interaction Dynamics through Distribution Matching Distillation in Few-Step Video World Models

Large video diffusion models offer expressive priors for embodied prediction and learning, yet their many-step sampling remains costly for interactive downstream use. Distribution Matching Distillation (DMD) enables few-step video generation, but can suppress robot--object motion while preserving visual quality. Examining DMD's teacher and fake-score signals, we find that weak re-noising keeps the teacher posterior concentrated near motion-deficient rollouts, limiting motion-restoring guidance. Meanwhile, stronger-motion rollouts tend to incur larger fake-score fitting errors, which can hinder the generator's learning of interaction dynamics. We propose DyMD, a DMD framework that adapts both teacher supervision and critic fitting to the evolving student. Temporal affinity--conditioned re-noise sampling adapts the timestep distribution to each rollout's current interaction fidelity by mixing the base schedule with a teacher prior motivated by local posterior variation, thereby balancing motion recovery and appearance refinement. To better track stronger-motion rollouts, dynamics-guided fake-score tracking uses a noise-conditioned predictor to estimate noise-relative fitting difficulty from latent temporal dynamics, then upweights predicted-hard rollouts in the critic loss. Using DyMD, we distill a 14B teacher into a four-step 1.3B student with no auxiliary modules at inference. On embodied-video benchmarks, the student improves R-Bench task adherence by $9.6$ percentage points and PAI-Bench-G Domain score by $5.1$ points over Base DMD while maintaining comparable visual quality. As a backbone for downstream action planning, our student achieves 34% mean success across two WorldArena tasks, compared with 16% for Base DMD.

cs.CV↗

arXiv2026

IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow

In image restoration, single-step discriminative mappings often lack fine details via expectation learning, whereas generative paradigms suffer from inefficient multi-step sampling and noise-residual coupling. To address this dilemma, we propose IR-Flow, a novel image restoration method based on Rectified Flow that serves as a unified framework bridging the gap between discriminative and generative paradigms. Specifically, we first construct multilevel data distribution flows, which expand the ability of models to learn from and adapt to various levels of degradation. Subsequently, cumulative velocity fields are proposed to learn transport trajectories across varying degradation levels, guiding intermediate states toward the clean target, while a multi-step consistency constraint is presented to enforce trajectory coherence and boost few-step restoration performance. We show that directly establishing a linear transport flow between degraded and clean image domains not only enables fast inference but also improves adaptability to out-of-distribution degradations. Extensive evaluations on deraining, denoising and raindrop removal tasks demonstrate that IR-Flow achieves competitive quantitative results with only a few sampling steps, offering an efficient and flexible framework that maintains an excellent distortion-perception balance. Our code is available at https://github.com/fanzh03/IR-Flow.

cs.CV↗

arXiv2026

Bird-SR: Bidirectional Reward-Guided Diffusion for Real-World Image Super-Resolution

Powered by multimodal text-to-image priors, diffusion-based super-resolution excels at synthesizing intricate details; however, models trained on synthetic low-resolution (LR) and high-resolution (HR) image pairs often degrade when applied to real-world LR images due to significant distribution shifts. We propose Bird-SR, a bidirectional reward-guided diffusion framework that formulates super-resolution as trajectory-level preference optimization via reward feedback learning (ReFL), jointly leveraging synthetic LR-HR pairs and real-world LR images. For structural fidelity easily affected in ReFL, the model is directly optimized on synthetic pairs at early diffusion steps, which also facilitates structure preservation for real-world inputs under smaller distribution gap in structure levels. For perceptual enhancement, quality-guided rewards are applied to both synthetic and real LR images at the later trajectory phase. To mitigate reward hacking, the rewards for synthetic results are formulated in a relative advantage space bounded by their ground-truth counterparts, while real-world optimization is regularized via a semantic alignment constraint. Furthermore, to balance structural and perceptual learning, we introduce a dynamic fidelity-perception weighting strategy that emphasizes structure preservation at early stages and progressively shifts focus toward perceptual optimization at later diffusion steps. Extensive experiments on real-world SR benchmarks demonstrate that Bird-SR consistently outperforms state-of-the-art methods in perceptual quality while preserving structural consistency, validating its effectiveness for real-world super-resolution. Our code can be obtained at https://github.com/fanzh03/Bird-SR.

cs.CV↗

arXiv2026

FinPercep-RM: A Fine-grained Reward Model and Co-evolutionary Curriculum for RL-based Real-world Super-Resolution

Reinforcement Learning with Human Feedback (RLHF) has proven effective in image generation field guided by reward models to align human preferences. Motivated by this, adapting RLHF for Image Super-Resolution (ISR) tasks has shown promise in optimizing perceptual quality with Image Quality Assessment (IQA) model as reward models. However, the traditional IQA model usually output a single global score, which are exceptionally insensitive to local and fine-grained distortions. This insensitivity allows ISR models to produce perceptually undesirable artifacts that yield spurious high scores, misaligning optimization objectives with perceptual quality and results in reward hacking. To address this, we propose a Fine-grained Perceptual Reward Model (FinPercep-RM) based on an Encoder-Decoder architecture. While providing a global quality score, it also generates a Perceptual Degradation Map that spatially localizes and quantifies local defects. We specifically introduce the FGR-30k dataset to train this model, consisting of diverse and subtle distortions from real-world super-resolution models. Despite the success of the FinPercep-RM model, its complexity introduces significant challenges in generator policy learning, leading to training instability. To address this, we propose a Co-evolutionary Curriculum Learning (CCL) mechanism, where both the reward model and the ISR model undergo synchronized curricula. The reward model progressively increases in complexity, while the ISR model starts with a simpler global reward for rapid convergence, gradually transitioning to the more complex model outputs. This easy-to-hard strategy enables stable training while suppressing reward hacking. Experiments validates the effectiveness of our method across ISR models in both global quality and local realism on RLHF methods.

cs.CV↗

arXiv2026

Semantic IDs for Recommender Systems at Snapchat: Use Cases, Technical Challenges, and Design Choices

Effective item identifiers (IDs) are an important component for recommender systems (RecSys) in practice, and are commonly adopted in many use cases such as retrieval and ranking. IDs can encode collaborative filtering signals within training data, such that RecSys models can extrapolate during the inference and personalize the prediction based on users' behavioral histories. Recently, Semantic IDs (SIDs) have become a trending paradigm for RecSys. In comparison to the conventional atomic ID, an SID is an ordered list of codes, derived from tokenizers such as residual quantization, applied to semantic representations commonly extracted from foundation models or collaborative signals. SIDs have drastically smaller cardinality than the atomic counterpart, and induce semantic clustering in the ID space. At Snapchat, we apply SIDs as auxiliary features for ranking models, and also explore SIDs as additional retrieval sources in different ML applications. In this paper, we discuss practical technical challenges we encountered while applying SIDs, experiments we have conducted, and design choices we have iterated to mitigate these challenges. Backed by promising offline results on both internal data and academic benchmarks as well as online A/B studies, SID variants have been launched in multiple production models with positive metrics impact.

cs.IR↗

arXiv2025

Elucidating and Endowing the Diffusion Training Paradigm for General Image Restoration

While diffusion models demonstrate strong generative capabilities in image restoration (IR) tasks, their complex architectures and iterative processes limit their practical application compared to mainstream reconstruction-based general ordinary IR networks. Existing approaches primarily focus on optimizing network architecture and diffusion paths but overlook the integration of the diffusion training paradigm within general ordinary IR frameworks. To address these challenges, this paper elucidates key principles for adapting the diffusion training paradigm to general IR training through systematic analysis of time-step dependencies, network hierarchies, noise-level relationships, and multi-restoration task correlations, proposing a new IR framework supported by diffusion-based training. To enable IR networks to simultaneously restore images and model generative representations, we introduce a series of regularization strategies that align diffusion objectives with IR tasks, improving generalization in single-task scenarios. Furthermore, recognizing that diffusion-based generation exerts varying influences across different IR tasks, we develop an incremental training paradigm and task-specific adaptors, further enhancing performance in multi-task unified IR. Experiments demonstrate that our method significantly improves the generalization of IR networks in single-task IR and achieves superior performance in multi-task unified IR. Notably, the proposed framework can be seamlessly integrated into existing general IR architectures.

cs.CV↗

arXiv2025

NTIRE 2025 Image Shadow Removal Challenge Report

This work examines the findings of the NTIRE 2025 Shadow Removal Challenge. A total of 306 participants have registered, with 17 teams successfully submitting their solutions during the final evaluation phase. Following the last two editions, this challenge had two evaluation tracks: one focusing on reconstruction fidelity and the other on visual perception through a user study. Both tracks were evaluated with images from the WSRD+ dataset, simulating interactions between self- and cast-shadows with a large number of diverse objects, textures, and materials.

cs.CV↗

arXiv2025

GiGL: Large-Scale Graph Neural Networks at Snapchat

Recent advances in graph machine learning (ML) with the introduction of Graph Neural Networks (GNNs) have led to a widespread interest in applying these approaches to business applications at scale. GNNs enable differentiable end-to-end (E2E) learning of model parameters given graph structure which enables optimization towards popular node, edge (link) and graph-level tasks. While the research innovation in new GNN layers and training strategies has been rapid, industrial adoption and utility of GNNs has lagged considerably due to the unique scale challenges that large-scale graph ML problems create. In this work, we share our approach to training, inference, and utilization of GNNs at Snapchat. To this end, we present GiGL (Gigantic Graph Learning), an open-source library to enable large-scale distributed graph ML to the benefit of researchers, ML engineers, and practitioners. We use GiGL internally at Snapchat to manage the heavy lifting of GNN workflows, including graph data preprocessing from relational DBs, subgraph sampling, distributed training, inference, and orchestration. GiGL is designed to interface cleanly with open-source GNN modeling libraries prominent in academia like PyTorch Geometric (PyG), while handling scaling and productionization challenges that make it easier for internal practitioners to focus on modeling. GiGL is used in multiple production settings, and has powered over 35 launches across multiple business domains in the last 2 years in the contexts of friend recommendation, content recommendation and advertising. This work details high-level design and tools the library provides, scaling properties, case studies in diverse business settings with industry-scale graphs, and several key lessons learned in employing graph ML at scale on large social data. GiGL is open-sourced at https://github.com/Snapchat/GiGL.

cs.LG↗

arXiv2024

Robust Training Objectives Improve Embedding-based Retrieval in Industrial Recommendation Systems

Improving recommendation systems (RS) can greatly enhance the user experience across many domains, such as social media. Many RS utilize embedding-based retrieval (EBR) approaches to retrieve candidates for recommendation. In an EBR system, the embedding quality is key. According to recent literature, self-supervised multitask learning (SSMTL) has showed strong performance on academic benchmarks in embedding learning and resulted in an overall improvement in multiple downstream tasks, demonstrating a larger resilience to the adverse conditions between each downstream task and thereby increased robustness and task generalization ability through the training objective. However, whether or not the success of SSMTL in academia as a robust training objectives translates to large-scale (i.e., over hundreds of million users and interactions in-between) industrial RS still requires verification. Simply adopting academic setups in industrial RS might entail two issues. Firstly, many self-supervised objectives require data augmentations (e.g., embedding masking/corruption) over a large portion of users and items, which is prohibitively expensive in industrial RS. Furthermore, some self-supervised objectives might not align with the recommendation task, which might lead to redundant computational overheads or negative transfer. In light of these two challenges, we evaluate using a robust training objective, specifically SSMTL, through a large-scale friend recommendation system on a social media platform in the tech sector, identifying whether this increase in robustness can work at scale in enhancing retrieval in the production setting. Through online A/B testing with SSMTL-based EBR, we observe statistically significant increases in key metrics in the friend recommendations, with up to 5.45% improvements in new friends made and 1.91% improvements in new friends made with cold-start users.

cs.IR↗

arXiv2018

Connectionist Recommendation in the Wild: On the utility and scrutability of neural networks for personalized course guidance

The aggregate behaviors of users can collectively encode deep semantic information about the objects with which they interact. In this paper, we demonstrate novel ways in which the synthesis of these data can illuminate the terrain of users' environment and support them in their decision making and wayfinding. A novel application of Recurrent Neural Networks and skip-gram models, approaches popularized by their application to modeling language, are brought to bear on student university enrollment sequences to create vector representations of courses and map out traversals across them. We present demonstrations of how scrutability from these neural networks can be gained and how the combination of these techniques can be seen as an evolution of content tagging and a means for a recommender to balance user preferences inferred from data with those explicitly specified. From validation of the models to the development of a UI, we discuss additional requisite functionality informed by the results of a usability study leading to the ultimate deployment of the system at a university.

cs.AI↗