Accurate reconstruction of long-duration neural recordings is challenging because local field potentials (LFPs) are high-resolution, multichannel, transient, and variable across subjects. We present PCA-DMD, a scalable operator-theoretic framework that segments LFP recordings into overlapping windows, projects them into a compact PCA space, learns linear Koopman evolution in the latent space, and reconstructs continuous signals through inverse projection and overlap-add aggregation. On 200,000-sample hippocampal recordings, PCA-DMD outperformed Classical DMD, SpDMD, MrDMD, and HODMD, achieving KLD=0.0761 and HD=0.0847. In all-pair cross-subject zero-shot generalization at 300,000 samples, correlations were 0.9504-0.9800, with HD=0.0010-0.0072 and KLD=0.0005-0.0022, without target-subject fine-tuning. Out-of-sample temporal prediction showed close one-step agreement on temporally held-out LFP segments across the unseen interval and multiple channels. Scalability analysis from 400,000 to 900,000 samples showed stable zero-shot reconstruction, with mean correlation remaining about 0.965-0.968 while computational cost increased predictably. External validation on an independent 93-channel Allen Neuropixels recording yielded mean and median channel-wise correlations of 0.7427 and 0.7990, respectively. Koopman spectral and mode analyses revealed dominant eigenvalues concentrated near the unit circle. PCA-DMD therefore provides an interpretable, generalizable, and computationally scalable framework for reconstructing high-dimensional neural dynamics.
展开完整摘要收起摘要↓
Accurate reconstruction of long-duration neural recordings is challenging because local field potentials (LFPs) are high-resolution, multichannel, transient, and variable across subjects. We present PCA-DMD, a scalable operator-theoretic framework that segments LFP recordings into overlapping windows, projects them into a compact PCA space, learns linear Koopman evolution in the latent space, and reconstructs continuous signals through inverse projection and overlap-add aggregation. On 200,000-sample hippocampal recordings, PCA-DMD outperformed Classical DMD, SpDMD, MrDMD, and HODMD, achieving KLD=0.0761 and HD=0.0847. In all-pair cross-subject zero-shot generalization at 300,000 samples, correlations were 0.9504-0.9800, with HD=0.0010-0.0072 and KLD=0.0005-0.0022, without target-subject fine-tuning. Out-of-sample temporal prediction showed close one-step agreement on temporally held-out LFP segments across the unseen interval and multiple channels. Scalability analysis from 400,000 to 900,000 samples showed stable zero-shot reconstruction, with mean correlation remaining about 0.965-0.968 while computational cost increased predictably. External validation on an independent 93-channel Allen Neuropixels recording yielded mean and median channel-wise correlations of 0.7427 and 0.7990, respectively. Koopman spectral and mode analyses revealed dominant eigenvalues concentrated near the unit circle. PCA-DMD therefore provides an interpretable, generalizable, and computationally scalable framework for reconstructing high-dimensional neural dynamics.
We study decentralized stochastic gradient tracking over a time-varying network of $N$ agents under a uniform window-mixing condition. Products of $τ$ consecutive doubly stochastic mixing matrices contract disagreement by a factor $λ<1$, although individual matrices need not contract disagreement strictly and individual communication graphs may be disconnected. We construct a time-varying quadratic norm that turns this window contraction into an exact one-step Lyapunov identity. This leads to coupled one-step recursions for the centroid and disagreement errors, without unrolling the dynamics over communication windows. For smooth strongly convex objectives, the leading stochastic term is $\widetilde{\mathcal O}(1/(NK))$; for smooth convex objectives, it is $\mathcal O(1/\sqrt{NK})$. Both match their centralized mini-batch counterparts and yield linear speedup after a network-dependent transient.
展开完整摘要收起摘要↓
We study decentralized stochastic gradient tracking over a time-varying network of $N$ agents under a uniform window-mixing condition. Products of $τ$ consecutive doubly stochastic mixing matrices contract disagreement by a factor $λ<1$, although individual matrices need not contract disagreement strictly and individual communication graphs may be disconnected. We construct a time-varying quadratic norm that turns this window contraction into an exact one-step Lyapunov identity. This leads to coupled one-step recursions for the centroid and disagreement errors, without unrolling the dynamics over communication windows. For smooth strongly convex objectives, the leading stochastic term is $\widetilde{\mathcal O}(1/(NK))$; for smooth convex objectives, it is $\mathcal O(1/\sqrt{NK})$. Both match their centralized mini-batch counterparts and yield linear speedup after a network-dependent transient.
作者Ao Shen, Yongheng Zhang, Yinghui Li, Manning Wang, Di Yin, Xing Sun
Large Multimodal Models (LMMs) for video reasoning have long been hindered by the high computational cost of processing vast amounts of visual information. This dilemma motivates the transfer of the reasoning capabilities of large models to smaller, more efficient ones. On-Policy Distillation (OPD) offers a promising solution by matching output-token distributions along student-generated trajectories. However, video reasoning often depends on evidence accumulated across multiple frames. In this context, output-level supervision only captures information expressed through token predictions and does not directly constrain the latent representations formed during reasoning. To address this limitation, we propose Latent-OPD, which augments OPD with trajectory-level latent distillation. Specifically, our method focuses on the position at the end of each trajectory, where hidden states effectively summarize the accumulated visual evidence and reasoning context. Furthermore, we introduce a progressive teacher-lookahead strategy, which aligns middle-to-late student layers with increasingly deeper teacher layers. Experiments on six video reasoning benchmarks show that Latent-OPD consistently outperforms output-only OPD. Notably, the improvements are particularly pronounced in scenarios with limited frames, long videos, or tasks requiring complex evidence aggregation. These results establish Latent-OPD as a highly effective approach to frame-efficient video reasoning.
展开完整摘要收起摘要↓
Large Multimodal Models (LMMs) for video reasoning have long been hindered by the high computational cost of processing vast amounts of visual information. This dilemma motivates the transfer of the reasoning capabilities of large models to smaller, more efficient ones. On-Policy Distillation (OPD) offers a promising solution by matching output-token distributions along student-generated trajectories. However, video reasoning often depends on evidence accumulated across multiple frames. In this context, output-level supervision only captures information expressed through token predictions and does not directly constrain the latent representations formed during reasoning. To address this limitation, we propose Latent-OPD, which augments OPD with trajectory-level latent distillation. Specifically, our method focuses on the position at the end of each trajectory, where hidden states effectively summarize the accumulated visual evidence and reasoning context. Furthermore, we introduce a progressive teacher-lookahead strategy, which aligns middle-to-late student layers with increasingly deeper teacher layers. Experiments on six video reasoning benchmarks show that Latent-OPD consistently outperforms output-only OPD. Notably, the improvements are particularly pronounced in scenarios with limited frames, long videos, or tasks requiring complex evidence aggregation. These results establish Latent-OPD as a highly effective approach to frame-efficient video reasoning.
作者Changhui Sun, Lanbo Liu, Hang Lei, Tong Ling, Jiahang Xie, Zhiyong Zheng, Yujia Wang, Hao Liu, Feng Xiao, Lu Liu, Yanlong Du, Zifeng Cheng, Ziwei Jiang, Qing Gu
On-policy distillation (OPD) aligns a student model with a teacher's logit distribution on student-generated trajectories. This approach has achieved strong empirical gains and can often surpass conventional off-policy distillation with substantially less data. However, standard token-level OPD can provide only fragmented corrections along an erroneous student trajectory and cannot unfold a complete and correct repair path. Motivated by this limitation, we propose Step-Level On-Policy Distillation (SOPD), which combines the long-horizon correction of supervised fine-tuning (SFT) with the on-policy advantage of OPD to provide step-level supervision over complete student-generated trajectories. We show that, at different limits of step length, SOPD reduces to SFT or approximates OPD. Compared with SFT, the teacher responses in SOPD are conditioned on student trajectories and therefore align more closely with student-visited states; compared with OPD, SOPD provides longer-horizon corrections rather than fragmented token-level guidance. Across both reasoning and agent tasks, SOPD substantially outperforms conventional SFT and OPD. For example, on ALFWorld, SOPD improves the average success rate by 13.4 points over Vanilla OPD. We hope this work offers a new perspective for future research on distillation methods.
展开完整摘要收起摘要↓
On-policy distillation (OPD) aligns a student model with a teacher's logit distribution on student-generated trajectories. This approach has achieved strong empirical gains and can often surpass conventional off-policy distillation with substantially less data. However, standard token-level OPD can provide only fragmented corrections along an erroneous student trajectory and cannot unfold a complete and correct repair path. Motivated by this limitation, we propose Step-Level On-Policy Distillation (SOPD), which combines the long-horizon correction of supervised fine-tuning (SFT) with the on-policy advantage of OPD to provide step-level supervision over complete student-generated trajectories. We show that, at different limits of step length, SOPD reduces to SFT or approximates OPD. Compared with SFT, the teacher responses in SOPD are conditioned on student trajectories and therefore align more closely with student-visited states; compared with OPD, SOPD provides longer-horizon corrections rather than fragmented token-level guidance. Across both reasoning and agent tasks, SOPD substantially outperforms conventional SFT and OPD. For example, on ALFWorld, SOPD improves the average success rate by 13.4 points over Vanilla OPD. We hope this work offers a new perspective for future research on distillation methods.
作者Zhaoyi Li, Deyang Kong, Yuan Wei, Evan Yang, Ranran Shen, Mahardika Krisna Ihsani, Ming Yang, Wei Zhang, Chuan Hao, Jian Yang, Ran Tao, Bryan Dai, Shikun Zhang, Wei Ye, Ying Wei, Defu Lian
On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.
展开完整摘要收起摘要↓
On-policy distillation (OPD) transfers teacher capabilities by supervising trajectories sampled from the student's own policy, yet its generalization behavior remains poorly understood, as most studies evaluate OPD on a single domain and on benchmarks close to the training data. We present a controlled study that varies one generalization factor at a time, from in-domain distribution shifts to cross-domain transfer and the multi-teacher setting. We find that OPD transfers a teacher's reasoning behavior rather than its answers to particular problems: training difficulty barely matters, and even problems the teacher never solves are useful. Transfer depends strongly on the origin relationship between teacher and student: same-origin pairs bring the student close to the teacher across languages, reasoning horizons, and even other domains, whereas cross-origin pairs mostly fit the trained distribution. This broad reach is a double-edged sword: since routing prompts to domain experts cannot confine each teacher's influence, combining them yields a mixture-dependent seesaw among their capabilities. These results clarify when OPD generalizes and offer a useful perspective for diagnosing multi-teacher OPD.
作者Keming Wu, Baoyi Wang, Kaichen Zhang, Xiang An, Zuhao Yang, Sudong Wang, Haowei Zhu, Tingxuan Huang, Hongcheng Gao, Bin Wang
Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to \textsc{StreamOPD}, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from $77.9%$ to $83.9%$---within $0.3$ points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by $9.1$ points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches $71.9%$ on OVO-Bench (excluding HLD) and $64.9%$ on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to $57.0%$, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.
展开完整摘要收起摘要↓
Streaming video understanding demands direct responses from the causally observed prefix of an unfolding video. Existing systems add inference-time memory, retrieval, and compression, yet a training-free sliding-window baseline already matches them. We therefore fix a memory-free recent-window protocol and ask how far post-training alone can go. Reinforcement learning with verifiable rewards fits this regime poorly, encouraging long ``think-then-answer'' generations, while on-policy distillation (OPD) supplies dense token-level teacher supervision on student trajectories but is stable only when both models train in thinking mode. These observations lead to \textsc{StreamOPD}, a recipe combining verifiable streaming-video data, thinking-mode OPD, and instruct-mode deployment. It raises StreamingBench from $77.9%$ to $83.9%$---within $0.3$ points of the 9B teacher---and improves OVO-Bench excluding its hallucination-detection subtask (HLD) by $9.1$ points under unchanged inference. As a teacher-privilege extension, Spatio-Temporal CueGate (ST-CueGate) aggregates cue-versus-no-cue teacher likelihood ratios into a group-relative response score that reweights OPD. It reaches $71.9%$ on OVO-Bench (excluding HLD) and $64.9%$ on Video-MME, and is the only variant that stays above the base model on all four benchmarks. Replacing the teacher with a frozen copy of the student's initial policy---on-policy self-distillation---retains most of these gains and lifts HLD to $57.0%$, above both the untrained student and the 9B teacher, so abstention loss is not intrinsic to the recipe. We provide a transparent and reproducible reference for open-source streaming-video research.
Long-horizon robot manipulation requires a robot to both execute individual skills reliably and sequence them coherently over extended tasks. Most hierarchical vision-language-action (VLA) models make each such decision with a single forward pass, leaving no mechanism to allocate additional computation to difficult or consequential choices. We introduce $τ_0$-VLA, a hierarchical robot foundation model that formulates high-level subtask generation as a compute-scalable inference problem through world-model-guided test-time computation. At each inference step, the high-level policy uses execution memory to generate a subtask and, when needed, searches over alternatives before committing to its output. A low-level policy then executes the generated subtask across multiple robot embodiments. The policy is trained on 40,115 hours of heterogeneous real-world data with multimodal co-training. Across in-domain and distribution-shifted settings, allocating additional test-time computation substantially improves next-subtask prediction accuracy, and these gains translate into higher closed-loop success on long-horizon robot manipulation tasks.
展开完整摘要收起摘要↓
Long-horizon robot manipulation requires a robot to both execute individual skills reliably and sequence them coherently over extended tasks. Most hierarchical vision-language-action (VLA) models make each such decision with a single forward pass, leaving no mechanism to allocate additional computation to difficult or consequential choices. We introduce $τ_0$-VLA, a hierarchical robot foundation model that formulates high-level subtask generation as a compute-scalable inference problem through world-model-guided test-time computation. At each inference step, the high-level policy uses execution memory to generate a subtask and, when needed, searches over alternatives before committing to its output. A low-level policy then executes the generated subtask across multiple robot embodiments. The policy is trained on 40,115 hours of heterogeneous real-world data with multimodal co-training. Across in-domain and distribution-shifted settings, allocating additional test-time computation substantially improves next-subtask prediction accuracy, and these gains translate into higher closed-loop success on long-horizon robot manipulation tasks.
Automated e-commerce poster design requires both high-quality poster generation and flexible editing of existing designs. However, most existing methods either target end-to-end poster generation or follow multi-stage design pipelines, with limited capability for flexible and precise editing of existing posters. To enable unified generation and editing of e-commerce posters, we introduce Text Patch Generation and Editing, a unified task formulation that treats text patches as atomic units and covers four operations: poster generation, patch addition, patch deletion, and patch modification, with optional reference-guided style control. Based on this, we propose PosterText, a unified model trained with a four-stage curriculum, including text rendering pretraining, instruction-following training, reinforcement learning for preference alignment, and spatial guidance self-distillation for execution refinement. We further construct a large-scale dataset with patch-level annotations and a comprehensive benchmark for evaluation. Extensive experiments demonstrate that PosterText achieves competitive performance against existing generation and editing approaches, validating the effectiveness of the proposed framework.
展开完整摘要收起摘要↓
Automated e-commerce poster design requires both high-quality poster generation and flexible editing of existing designs. However, most existing methods either target end-to-end poster generation or follow multi-stage design pipelines, with limited capability for flexible and precise editing of existing posters. To enable unified generation and editing of e-commerce posters, we introduce Text Patch Generation and Editing, a unified task formulation that treats text patches as atomic units and covers four operations: poster generation, patch addition, patch deletion, and patch modification, with optional reference-guided style control. Based on this, we propose PosterText, a unified model trained with a four-stage curriculum, including text rendering pretraining, instruction-following training, reinforcement learning for preference alignment, and spatial guidance self-distillation for execution refinement. We further construct a large-scale dataset with patch-level annotations and a comprehensive benchmark for evaluation. Extensive experiments demonstrate that PosterText achieves competitive performance against existing generation and editing approaches, validating the effectiveness of the proposed framework.
作者Leon Todorov, Peter Rot, Peter Peer, Vitomir Štruc, Klemen Grm
Face recognition templates are compact identity representations, yet they also encode rich semantic information about facial appearance. Prior work has shown that templates can be inverted to images or indirectly manipulated through image-editing pipelines, but direct semantic editing in template space remains largely unexplored. Existing interpretability methods for face recognition often rely on manual neuron inspection or predefined attribute labels, limiting scalability and semantic flexibility. To address this gap, we propose SCOUT (Semantic Concept Discovery for Open-VocabUlary Editing of Face Recognition Templates), an end-to-end framework for discovering and directly manipulating semantic concepts in face recognition templates using mechanistic interpretability. SCOUT learns sparse template representations, generates semantic hypotheses for latent features from natural-language descriptions, and validates their stability. The resulting features act as controllable semantic directions for direct editing, avoiding costly edit--re-encode pipelines. Experiments with face recognition models using CNN, ViT, and Swin backbones show that SCOUT discovers interpretable concepts beyond standard attribute labels and enables controllable, identity-aware template manipulation with negligible impact on identity matching. We further show that edited templates can subsequently be decoded with independent inversion models for visualization and evaluation.
展开完整摘要收起摘要↓
Face recognition templates are compact identity representations, yet they also encode rich semantic information about facial appearance. Prior work has shown that templates can be inverted to images or indirectly manipulated through image-editing pipelines, but direct semantic editing in template space remains largely unexplored. Existing interpretability methods for face recognition often rely on manual neuron inspection or predefined attribute labels, limiting scalability and semantic flexibility. To address this gap, we propose SCOUT (Semantic Concept Discovery for Open-VocabUlary Editing of Face Recognition Templates), an end-to-end framework for discovering and directly manipulating semantic concepts in face recognition templates using mechanistic interpretability. SCOUT learns sparse template representations, generates semantic hypotheses for latent features from natural-language descriptions, and validates their stability. The resulting features act as controllable semantic directions for direct editing, avoiding costly edit--re-encode pipelines. Experiments with face recognition models using CNN, ViT, and Swin backbones show that SCOUT discovers interpretable concepts beyond standard attribute labels and enables controllable, identity-aware template manipulation with negligible impact on identity matching. We further show that edited templates can subsequently be decoded with independent inversion models for visualization and evaluation.
Meal awareness can help people reflect on hydration, chewing rhythm, and conversation-heavy meals, but many eating-sensing approaches rely on cameras, microphones, food photographs, or repeated self-logging. PPG and IMU offer a narrower sensing path by capturing physiological and motion patterns around meal-adjacent actions without raw audio, video, or photographs. We present Oto-Meal, an audio- and image-free earable prototype. Its pooled neural recognizer uses a two-stage event/rest gate and five-class behavior classifier. Separately, a within-user protocol evaluates a lightweight memory matcher built from labeled target-user examples. We invited seven volunteers and collected a seven-user dataset for mixed-user training, within-user memory evaluation, and modality ablation. The pooled model reaches 70.99% event accuracy. Under the separate memory protocol, 20% target-user calibration reaches 80.38 $\pm$ 0.84% event accuracy and 81.77 $\pm$ 0.69% cascade accuracy; with 60% calibration, PPG+IMU reaches 85.13 $\pm$ 0.57% event accuracy and outperforms IMU-only and PPG-only. These preliminary results suggest that audio- and image-free earable sensing with inspectable personalization can support low-burden meal-awareness review.
展开完整摘要收起摘要↓
Meal awareness can help people reflect on hydration, chewing rhythm, and conversation-heavy meals, but many eating-sensing approaches rely on cameras, microphones, food photographs, or repeated self-logging. PPG and IMU offer a narrower sensing path by capturing physiological and motion patterns around meal-adjacent actions without raw audio, video, or photographs. We present Oto-Meal, an audio- and image-free earable prototype. Its pooled neural recognizer uses a two-stage event/rest gate and five-class behavior classifier. Separately, a within-user protocol evaluates a lightweight memory matcher built from labeled target-user examples. We invited seven volunteers and collected a seven-user dataset for mixed-user training, within-user memory evaluation, and modality ablation. The pooled model reaches 70.99% event accuracy. Under the separate memory protocol, 20% target-user calibration reaches 80.38 $\pm$ 0.84% event accuracy and 81.77 $\pm$ 0.69% cascade accuracy; with 60% calibration, PPG+IMU reaches 85.13 $\pm$ 0.57% event accuracy and outperforms IMU-only and PPG-only. These preliminary results suggest that audio- and image-free earable sensing with inspectable personalization can support low-burden meal-awareness review.
作者Yuji Wang, Yuheng Chen, Teng Hu, Ran Yi, Yijia Hong, Han Feng, Weijian Cao, Chengjie Wang, Lizhuang Ma, Jiangning Zhang
Video generation is rapidly evolving from single-shot clips to multi-shot narratives, where the human character serves as the core narrative anchor. However, existing benchmarks mainly assess character appearance or individual-shot quality, without measuring whether physical and emotional states remain coherent across cuts. They also rarely provide criterion-specific evaluation methods, although physical continuity, facial dynamics, and cinematic relations require different visual, temporal, and relational evidence. To address these limitations, we introduce PersonaShot, the first person-centric benchmark for narrative continuity in multi-shot video generation. PersonaShot contains approximately 1,000 multi-shot segments and 16 metrics spanning physical continuity, affective dynamics, and cinematic grammar. \textbf{1) Narrative Continuity Benchmark:} We evaluate character coherence across three temporal levels: within-shot states, cross-shot transitions, and sequence-level trajectories. \textbf{2) Human-Aligned Specialist Evaluators:} We distill reasoning from a large multimodal teacher into lightweight criterion-specific evaluators, each grounded in the visual, temporal, or relational evidence required by its metric, and align them with expert human judgments. \textbf{3) Systematic Evaluation and Insights:} Our evaluation reveals distinct capability profiles across state-of-the-art models and a clear gap between perceptual quality and cross-shot narrative continuity. Even visually compelling videos frequently exhibit physical-state resets, abrupt affective shifts, and broken cinematic relations across shots. Human studies further demonstrate strong agreement between our evaluators and expert judgments.
展开完整摘要收起摘要↓
Video generation is rapidly evolving from single-shot clips to multi-shot narratives, where the human character serves as the core narrative anchor. However, existing benchmarks mainly assess character appearance or individual-shot quality, without measuring whether physical and emotional states remain coherent across cuts. They also rarely provide criterion-specific evaluation methods, although physical continuity, facial dynamics, and cinematic relations require different visual, temporal, and relational evidence. To address these limitations, we introduce PersonaShot, the first person-centric benchmark for narrative continuity in multi-shot video generation. PersonaShot contains approximately 1,000 multi-shot segments and 16 metrics spanning physical continuity, affective dynamics, and cinematic grammar. \textbf{1) Narrative Continuity Benchmark:} We evaluate character coherence across three temporal levels: within-shot states, cross-shot transitions, and sequence-level trajectories. \textbf{2) Human-Aligned Specialist Evaluators:} We distill reasoning from a large multimodal teacher into lightweight criterion-specific evaluators, each grounded in the visual, temporal, or relational evidence required by its metric, and align them with expert human judgments. \textbf{3) Systematic Evaluation and Insights:} Our evaluation reveals distinct capability profiles across state-of-the-art models and a clear gap between perceptual quality and cross-shot narrative continuity. Even visually compelling videos frequently exhibit physical-state resets, abrupt affective shifts, and broken cinematic relations across shots. Human studies further demonstrate strong agreement between our evaluators and expert judgments.
Cross-border e-commerce image translation is essential for global retail, where product images, banners, and detail pages need to be produced in different languages. Existing methods struggle to achieve accurate translation, faithful visual identity preservation, and easy-to-edit outputs, simultaneously. To address these challenges, we introduce TransAnyText, a structured visual code framework that reformulates image text translation as generating renderable HTML patches from source images and target languages. Our framework decouples semantic generation from pixel rendering: a vision-language model (VLM) handles visual understanding, cross-lingual translation, and structured visual generation, while a diffusion model performs background inpainting and pixel-level refinement, followed by deterministic rendering to synthesize the final image. Based on this formulation, we develop a three-stage post-training framework, where supervised fine-tuning (SFT) establishes the image-to-code mapping, privilege-gap weighted self-distillation (PWSD) improves the learning of style and layout tokens, and reinforcement learning with verifiable rewards (RLVR) further optimizes task-level performance. We further introduce TransAnyDataset and TransAnyBench, a multilingual dataset and benchmark for e-commerce image translation. Extensive experiments demonstrate competitive performance against cascaded pipelines, open-source end-to-end models, and closed-source image editing systems, providing an effective, controllable, and editable solution for cross-border e-commerce image translation.
展开完整摘要收起摘要↓
Cross-border e-commerce image translation is essential for global retail, where product images, banners, and detail pages need to be produced in different languages. Existing methods struggle to achieve accurate translation, faithful visual identity preservation, and easy-to-edit outputs, simultaneously. To address these challenges, we introduce TransAnyText, a structured visual code framework that reformulates image text translation as generating renderable HTML patches from source images and target languages. Our framework decouples semantic generation from pixel rendering: a vision-language model (VLM) handles visual understanding, cross-lingual translation, and structured visual generation, while a diffusion model performs background inpainting and pixel-level refinement, followed by deterministic rendering to synthesize the final image. Based on this formulation, we develop a three-stage post-training framework, where supervised fine-tuning (SFT) establishes the image-to-code mapping, privilege-gap weighted self-distillation (PWSD) improves the learning of style and layout tokens, and reinforcement learning with verifiable rewards (RLVR) further optimizes task-level performance. We further introduce TransAnyDataset and TransAnyBench, a multilingual dataset and benchmark for e-commerce image translation. Extensive experiments demonstrate competitive performance against cascaded pipelines, open-source end-to-end models, and closed-source image editing systems, providing an effective, controllable, and editable solution for cross-border e-commerce image translation.
作者Yuji Wang, Teng Hu, Yuheng Chen, Ran Yi, Han Feng, Weijian Cao, Chengjie Wang, Lizhuang Ma, Jiangning Zhang
Despite progress in instruction-based video editing, unimodal textual instructions inherently struggle to convey fine-grained textures and complex dynamics. To bridge this perceptual gap, we propose Visual In-context Editing, a new paradigm elevating video editing from textual instructions to multi-modal visual guidance encompassing single image, image pair, and video pair. To facilitate this paradigm, we curate VicEdit-400K, the first large-scale dataset for visual in-context video editing. We develop an automated pipeline to generate 400K high-quality samples across ten task types, ensuring superior visual fidelity and semantic consistency through multi-dimensional filtering. Leveraging this foundation, we introduce VicEdit, a unified framework to bridge visual and textual contexts. To adaptively extract editing semantics from heterogeneous references, we design Modality-Adaptive Semantic Distillation, which produces modality-specific semantic tokens from visual references. These tokens are then synergistically integrated with textual instructions through Dual-Context Injection, enabling the generation process to benefit from both visual and textual signals. Extensive evaluations on VicEditBench demonstrate that VicEdit achieves state-of-the-art performance across both basic instruction editing and visual in-context editing tasks, establishing visual in-context learning as a powerful and controllable paradigm for video editing.
展开完整摘要收起摘要↓
Despite progress in instruction-based video editing, unimodal textual instructions inherently struggle to convey fine-grained textures and complex dynamics. To bridge this perceptual gap, we propose Visual In-context Editing, a new paradigm elevating video editing from textual instructions to multi-modal visual guidance encompassing single image, image pair, and video pair. To facilitate this paradigm, we curate VicEdit-400K, the first large-scale dataset for visual in-context video editing. We develop an automated pipeline to generate 400K high-quality samples across ten task types, ensuring superior visual fidelity and semantic consistency through multi-dimensional filtering. Leveraging this foundation, we introduce VicEdit, a unified framework to bridge visual and textual contexts. To adaptively extract editing semantics from heterogeneous references, we design Modality-Adaptive Semantic Distillation, which produces modality-specific semantic tokens from visual references. These tokens are then synergistically integrated with textual instructions through Dual-Context Injection, enabling the generation process to benefit from both visual and textual signals. Extensive evaluations on VicEditBench demonstrate that VicEdit achieves state-of-the-art performance across both basic instruction editing and visual in-context editing tasks, establishing visual in-context learning as a powerful and controllable paradigm for video editing.
作者Long Cui, Xiaoqian Liu, Qi Qin, Yi Xin, Tao Lin, Jianguo Li, Linfeng Zhang
Existing image editing frameworks predominantly follow the training paradigm of text-to-image diffusion models. However, extending this paradigm to image editing highlights two inherent discrepancies, specifically, the insufficient attention to edit concept granularity and the training inefficiency caused by sparse supervision signals. To address these issues, we establish a comprehensive hierarchical taxonomy featuring over 1,000 fine-grained edit concepts and build ConceptEdit-12M, a massive dataset of 12 million high-quality editing pairs via an improved synthesis framework. This library-driven approach effectively rectifies the distribution collapse of generated data while ensuring high data fidelity. Furthermore, we propose a dense supervision training strategy that synthesizes multiple non-interfering concepts into single image pairs. By providing richer learning signals, this strategy significantly enhances both training efficiency and overall model performance. Training results validate our strategy, significantly outperforming prior works. Finally, we present ConceptEdit-Bench, a granular evaluation suite designed to diagnose model capabilities across a vast array of real-world scenarios.
展开完整摘要收起摘要↓
Existing image editing frameworks predominantly follow the training paradigm of text-to-image diffusion models. However, extending this paradigm to image editing highlights two inherent discrepancies, specifically, the insufficient attention to edit concept granularity and the training inefficiency caused by sparse supervision signals. To address these issues, we establish a comprehensive hierarchical taxonomy featuring over 1,000 fine-grained edit concepts and build ConceptEdit-12M, a massive dataset of 12 million high-quality editing pairs via an improved synthesis framework. This library-driven approach effectively rectifies the distribution collapse of generated data while ensuring high data fidelity. Furthermore, we propose a dense supervision training strategy that synthesizes multiple non-interfering concepts into single image pairs. By providing richer learning signals, this strategy significantly enhances both training efficiency and overall model performance. Training results validate our strategy, significantly outperforming prior works. Finally, we present ConceptEdit-Bench, a granular evaluation suite designed to diagnose model capabilities across a vast array of real-world scenarios.
We present AnyTalk, a novel method for generating 3D speech animations for arbitrary characters without requiring any animation data. While existing audio-driven 3D speech animation methods rely on character-specific training data or laborious rigging/re-meshing, AnyTalk circumvents these limitations by leveraging recent video diffusion models trained on extensive video datasets. We first adapt a pre-trained video diffusion model to a target character through our Character-specific Fine-tuning (CsF) technique. By fine-tuning on rendered images of the 3D character paired with zeroed-out audio embeddings (representing "no motion"), we eliminate the need for animation data while preserving the motion prior of large-scale video diffusion model. We then uplift the resulting talking-head video into a 3D speech animation by estimating blendshape parameters through a proposed optimization process. AnyTalk enables lip-synced animations across diverse face meshes and blendshape configurations, significantly reducing manual effort and data requirements. We further enhance usability by distilling AnyTalk into a streamlined network, $\text{AnyTalk}_{RT}$, thereby enabling real-time performance. By leveraging talking-head video generation, our method broadens access to audio-driven speech animation technology for arbitrary characters. The code is publicly available at https://serin-yoon.github.io/projects/anytalk/.
展开完整摘要收起摘要↓
We present AnyTalk, a novel method for generating 3D speech animations for arbitrary characters without requiring any animation data. While existing audio-driven 3D speech animation methods rely on character-specific training data or laborious rigging/re-meshing, AnyTalk circumvents these limitations by leveraging recent video diffusion models trained on extensive video datasets. We first adapt a pre-trained video diffusion model to a target character through our Character-specific Fine-tuning (CsF) technique. By fine-tuning on rendered images of the 3D character paired with zeroed-out audio embeddings (representing "no motion"), we eliminate the need for animation data while preserving the motion prior of large-scale video diffusion model. We then uplift the resulting talking-head video into a 3D speech animation by estimating blendshape parameters through a proposed optimization process. AnyTalk enables lip-synced animations across diverse face meshes and blendshape configurations, significantly reducing manual effort and data requirements. We further enhance usability by distilling AnyTalk into a streamlined network, $\text{AnyTalk}_{RT}$, thereby enabling real-time performance. By leveraging talking-head video generation, our method broadens access to audio-driven speech animation technology for arbitrary characters. The code is publicly available at https://serin-yoon.github.io/projects/anytalk/.
Unified image restoration (UIR) aims to recover high-quality (HQ) content from low-quality (LQ) images with different degradations using a single model. Most recent methods adapt large pretrained text-to-image (T2I) latent diffusion models for their strong capacity and generative priors. However, the variational autoencoder (VAE) in latent T2I models may discard restoration-sensitive details, while the open-ended synthesis prior can introduce content-inconsistent artifacts. We present PixRestore, a VAE-free pixel-space Diffusion Transformer (DiT) for UIR, where the diffusion backbone is trained entirely from scratch, without relying on T2I pretraining. PixRestore performs flow matching directly on patchified pixels, preserving fine-grained details while keeping the token sequence tractable. To adapt to different degradations, PixRestore learns to predict the reliability of layer features using LQ--HQ DINO feature similarity. Features from more reliable layers are fused as dense conditioning, while less reliable layers receive stronger HQ-feature supervision to encourage degradation removal. We train PixRestore on a large-scale corpus of diverse scenes and degradations, and further finetune it into a one-step generator using DINO-based adversarial objectives for efficient inference. Experiments on public benchmarks and real-world test sets show that, with only about 50M parameters and single-step inference, PixRestore achieves the best overall fidelity, perceptual quality, and robustness to degradations among competing UIR models while being far more efficient. Larger PixRestore variants can further boost performance, demonstrating the scalability of our pixel-space design. Code and the curated benchmark can be found at https://github.com/csslc/PixRestore.
展开完整摘要收起摘要↓
Unified image restoration (UIR) aims to recover high-quality (HQ) content from low-quality (LQ) images with different degradations using a single model. Most recent methods adapt large pretrained text-to-image (T2I) latent diffusion models for their strong capacity and generative priors. However, the variational autoencoder (VAE) in latent T2I models may discard restoration-sensitive details, while the open-ended synthesis prior can introduce content-inconsistent artifacts. We present PixRestore, a VAE-free pixel-space Diffusion Transformer (DiT) for UIR, where the diffusion backbone is trained entirely from scratch, without relying on T2I pretraining. PixRestore performs flow matching directly on patchified pixels, preserving fine-grained details while keeping the token sequence tractable. To adapt to different degradations, PixRestore learns to predict the reliability of layer features using LQ--HQ DINO feature similarity. Features from more reliable layers are fused as dense conditioning, while less reliable layers receive stronger HQ-feature supervision to encourage degradation removal. We train PixRestore on a large-scale corpus of diverse scenes and degradations, and further finetune it into a one-step generator using DINO-based adversarial objectives for efficient inference. Experiments on public benchmarks and real-world test sets show that, with only about 50M parameters and single-step inference, PixRestore achieves the best overall fidelity, perceptual quality, and robustness to degradations among competing UIR models while being far more efficient. Larger PixRestore variants can further boost performance, demonstrating the scalability of our pixel-space design. Code and the curated benchmark can be found at https://github.com/csslc/PixRestore.
Despite recent advances in unified multimodal models for multi-reference image generation, existing benchmarks remain organized around predefined task types (e.g., "subject composition"), which are ill-suited to this combinatorial setting and lead to fragmented coverage, uncontrolled complexity, and little diagnostic value. Recognizing that diverse multi-reference tasks share a common set of atomic operations, we adopt a capability-oriented perspective and formalize four operators: Anchor ($f$), Disentangle ($g$), Apply ($\oplus$), and Compose ($C$). Any multi-reference prompt can then be represented as a compositional formula over these operators, whose structural complexity is quantified by the number of operator slots. Building on this formulation, we construct TRACE-Bench, comprising approximately 1,600 evaluation cases across slot counts 1--8, built from 631 formula templates and around 4,000 reference images spanning diverse artistic styles and real-world subjects. The formula structure directly drives an operator-aligned evaluation protocol for per-capability scoring and a diagnostic tree analysis for recursive failure localization. Evaluating 9 leading models reveals insights invisible to holistic scoring: the primary bottleneck lies in disentanglement ($g$) and attribute binding ($\oplus$) rather than scene-level composition ($C$), with even the best model scoring only 0.74 on attribute fidelity. Project page: https://amuseum-whr.github.io/TraceBench
展开完整摘要收起摘要↓
Despite recent advances in unified multimodal models for multi-reference image generation, existing benchmarks remain organized around predefined task types (e.g., "subject composition"), which are ill-suited to this combinatorial setting and lead to fragmented coverage, uncontrolled complexity, and little diagnostic value. Recognizing that diverse multi-reference tasks share a common set of atomic operations, we adopt a capability-oriented perspective and formalize four operators: Anchor ($f$), Disentangle ($g$), Apply ($\oplus$), and Compose ($C$). Any multi-reference prompt can then be represented as a compositional formula over these operators, whose structural complexity is quantified by the number of operator slots. Building on this formulation, we construct TRACE-Bench, comprising approximately 1,600 evaluation cases across slot counts 1--8, built from 631 formula templates and around 4,000 reference images spanning diverse artistic styles and real-world subjects. The formula structure directly drives an operator-aligned evaluation protocol for per-capability scoring and a diagnostic tree analysis for recursive failure localization. Evaluating 9 leading models reveals insights invisible to holistic scoring: the primary bottleneck lies in disentanglement ($g$) and attribute binding ($\oplus$) rather than scene-level composition ($C$), with even the best model scoring only 0.74 on attribute fidelity. Project page: https://amuseum-whr.github.io/TraceBench
Instruction-based general video editing seeks to unify diverse editing operations within a single, intuitive interface. Existing approaches often rely on resource-intensive conditioning, using either heavyweight branches or costly source concatenation. Is there any efficient way to model editing intent? Thus, we introduce GRNEdit, a lightweight two-stage framework. GRN inspires our approach by encoding visual semantics through combinations of bits. Through task-specific fine-tuning, we take this representation further and recast editing semantics as local retain-or-flip decisions over individual bits. Source information is consequently modeled as coordinate-wise evidence supporting the observed binary states, while the GRN backbone remains responsible for resolving their global composition into coherent generative semantics. In Stage I, a compact encoder translates discrete source codes into continuous evidence signals, which GRN assimilates throughout binary refinement. Inspired by null-prompt training for classifier-free guidance, we further assign the null condition an editing-specific meaning: an empty instruction denotes no edit and is supervised through source reconstruction. This identity pathway not only implicitly strengthens evidence utilization and content preservation in Stage I, but also produces a source-preserving state in the same representation space as the edited state. Stage II can therefore directly compare each edited state with its source-preserving counterpart and use their discrepancy to revise unresolved target-bit decisions. Trained on only 0.6M pairs with less than 3% conditioning parameters, GRNEdit-2B and GRNEdit-8B achieve scores of 4.03 and 4.18 on OpenVE-Bench. The 2B model outperforms multiple 14B open-source editors, while the 8B model performs on par with leading open-source editors.
展开完整摘要收起摘要↓
Instruction-based general video editing seeks to unify diverse editing operations within a single, intuitive interface. Existing approaches often rely on resource-intensive conditioning, using either heavyweight branches or costly source concatenation. Is there any efficient way to model editing intent? Thus, we introduce GRNEdit, a lightweight two-stage framework. GRN inspires our approach by encoding visual semantics through combinations of bits. Through task-specific fine-tuning, we take this representation further and recast editing semantics as local retain-or-flip decisions over individual bits. Source information is consequently modeled as coordinate-wise evidence supporting the observed binary states, while the GRN backbone remains responsible for resolving their global composition into coherent generative semantics. In Stage I, a compact encoder translates discrete source codes into continuous evidence signals, which GRN assimilates throughout binary refinement. Inspired by null-prompt training for classifier-free guidance, we further assign the null condition an editing-specific meaning: an empty instruction denotes no edit and is supervised through source reconstruction. This identity pathway not only implicitly strengthens evidence utilization and content preservation in Stage I, but also produces a source-preserving state in the same representation space as the edited state. Stage II can therefore directly compare each edited state with its source-preserving counterpart and use their discrepancy to revise unresolved target-bit decisions. Trained on only 0.6M pairs with less than 3% conditioning parameters, GRNEdit-2B and GRNEdit-8B achieve scores of 4.03 and 4.18 on OpenVE-Bench. The 2B model outperforms multiple 14B open-source editors, while the 8B model performs on par with leading open-source editors.
Video Diffusion Transformers (DiTs) spend most of their compute inside the Self-Attention operation, whose cost grows quadratically, $\mathcal{O}(n^2)$, with the number of latent tokens $n$. For the task of video generation, the token count is large, so this term dominates runtime and memory, and thereby caps the resolution and duration we can generate. Linear $\mathcal{O}(n)$ and low-rank $\mathcal{O}(nk)$ surrogates of Self-Attention trade the full softmax $QK^T$ for cheaper kernels, but rarely recover the original's expressivity, leaving a stubborn quality gap. Motivated by this, we propose SQuad, a Sub-Quadratic Attention Distillation framework that achieves a complexity of $\mathcal{O}(n\sqrt{n})$ in the resulting distilled Attention, naturally balancing the efficiency v/s expressivity trade-off. Instead of training our own Video DiT from scratch, which is prohibitively expensive, we fit a pretrained full softmax Self-Attention DiT into our proposed SQuad-Attention one by distilling the former in two stages: Flow-Matching Supervised Fine-Tuning (SFT), followed by improved Distribution Matching Distillation (DMD2) which additionally makes the sampling more efficient. On the Wan~2.2 5B text-to-video model, SQuAD matches the quadratic teacher on VBench ($83.20$ v/s $83.08$) while cutting the per-step per-block attention FLOPs by $\sim$$67\times$ and attention latency by $\sim$$11\times$, and end-to-end DiT latency by 2$\times$, all while also generating a video in only $6$ Neural Functional Evaluations (NFEs) instead of the default $100$.
展开完整摘要收起摘要↓
Video Diffusion Transformers (DiTs) spend most of their compute inside the Self-Attention operation, whose cost grows quadratically, $\mathcal{O}(n^2)$, with the number of latent tokens $n$. For the task of video generation, the token count is large, so this term dominates runtime and memory, and thereby caps the resolution and duration we can generate. Linear $\mathcal{O}(n)$ and low-rank $\mathcal{O}(nk)$ surrogates of Self-Attention trade the full softmax $QK^T$ for cheaper kernels, but rarely recover the original's expressivity, leaving a stubborn quality gap. Motivated by this, we propose SQuad, a Sub-Quadratic Attention Distillation framework that achieves a complexity of $\mathcal{O}(n\sqrt{n})$ in the resulting distilled Attention, naturally balancing the efficiency v/s expressivity trade-off. Instead of training our own Video DiT from scratch, which is prohibitively expensive, we fit a pretrained full softmax Self-Attention DiT into our proposed SQuad-Attention one by distilling the former in two stages: Flow-Matching Supervised Fine-Tuning (SFT), followed by improved Distribution Matching Distillation (DMD2) which additionally makes the sampling more efficient. On the Wan~2.2 5B text-to-video model, SQuAD matches the quadratic teacher on VBench ($83.20$ v/s $83.08$) while cutting the per-step per-block attention FLOPs by $\sim$$67\times$ and attention latency by $\sim$$11\times$, and end-to-end DiT latency by 2$\times$, all while also generating a video in only $6$ Neural Functional Evaluations (NFEs) instead of the default $100$.
Video world models approximate the stochastic distribution of physical outcomes through generative sampling, but existing benchmarks score individual generations or compare distributions coarsely over a whole dataset, leaving the fine-grained aleatoric uncertainty of specific phenomena untested. We introduce CaliBench, which scores outcomes in a physically interpretable discrete space - a bin index, a die face, a suit, a colour - rather than a learned feature space such as in FID, so the distance from a known reference distribution is measured directly. We curate outcome spaces whose reference is known in closed form (binomial Galton boards, Bernoulli forks, uniform dice/cards/lottery, a skewed European-roulette colour), enabling an exact calibration test. We decompose performance into two orthogonal axes that a single accuracy metric conflates: scorability, the fraction of generations yielding a scoreable outcome, and calibration, the total variation distance from the reference on that sample. A chi-squared test assesses significance; as calibration is its null hypothesis it can evidence only miscalibration, and at N=32 per cell detects only large deviations. We apply it to nine scenes and six image-to-video models (WAN-2.7, SeeDance-2.0, HappyHorse-1.0, Veo 3.1, Runway Gen-4.5, Cosmos3-Super), 32 generations each. Models consistently concentrate probability mass on a few outcomes rather than reproducing the reference. Most scene-model combinations are significantly miscalibrated, in the extreme collapsing to one outcome, as Veo 3.1 does on dice. On roulette, generations often leave the ball ambiguously placed, giving several models low scorability. Performance varies by scene: no model dominates all nine. We release the protocol and a metric (mean normalised total variation, mnTV) for comparing new models against our results.
展开完整摘要收起摘要↓
Video world models approximate the stochastic distribution of physical outcomes through generative sampling, but existing benchmarks score individual generations or compare distributions coarsely over a whole dataset, leaving the fine-grained aleatoric uncertainty of specific phenomena untested. We introduce CaliBench, which scores outcomes in a physically interpretable discrete space - a bin index, a die face, a suit, a colour - rather than a learned feature space such as in FID, so the distance from a known reference distribution is measured directly. We curate outcome spaces whose reference is known in closed form (binomial Galton boards, Bernoulli forks, uniform dice/cards/lottery, a skewed European-roulette colour), enabling an exact calibration test. We decompose performance into two orthogonal axes that a single accuracy metric conflates: scorability, the fraction of generations yielding a scoreable outcome, and calibration, the total variation distance from the reference on that sample. A chi-squared test assesses significance; as calibration is its null hypothesis it can evidence only miscalibration, and at N=32 per cell detects only large deviations. We apply it to nine scenes and six image-to-video models (WAN-2.7, SeeDance-2.0, HappyHorse-1.0, Veo 3.1, Runway Gen-4.5, Cosmos3-Super), 32 generations each. Models consistently concentrate probability mass on a few outcomes rather than reproducing the reference. Most scene-model combinations are significantly miscalibrated, in the extreme collapsing to one outcome, as Veo 3.1 does on dice. On roulette, generations often leave the ball ambiguously placed, giving several models low scorability. Performance varies by scene: no model dominates all nine. We release the protocol and a metric (mean normalised total variation, mnTV) for comparing new models against our results.
作者Jianjie Luo, Yiming Zhong, Haoming Shen, Yupeng Xiao, Zhenguo Yang
Identity-preserving video generation (IPVG) requires synthesizing videos that are faithful to both reference subjects and text prompts. Existing methods are often hindered by high tuning costs or limited input-level enhancements, struggling to maintain rigid identity consistency during complex, long-sequence actions. To address these limitations, we propose KeyID, a training-free IPVG framework that decouples the synthesis of video dynamics from the injection of identity. Specifically, KeyID comprises two components: (1) Reference-Aware Video Generation, which produces an identity-agnostic video draft aligned with multiple references, and (2) Identity-Preserved Keyframe Editing, which integrates the target identity via sparse keyframe correction and subsequent motion interpolation. By shifting from dense frame-level supervision to sparse keyframe-level refinement, KeyID effectively resolves the capacity conflict between prompt adherence and identity fidelity. Crucially, our modular design allows seamless extension to multi-subject references and complex sequential action generation without additional training. KeyID outperforms prior works and is validated by automatic and human evaluations on the official challenge benchmark, ultimately securing the runner-up position in the Track 2 (Sequential Action) of the ACM Multimedia 2026 IPVG Grand Challenge. Source code is available at https://github.com/WISLab-GDUT/KeyID.
展开完整摘要收起摘要↓
Identity-preserving video generation (IPVG) requires synthesizing videos that are faithful to both reference subjects and text prompts. Existing methods are often hindered by high tuning costs or limited input-level enhancements, struggling to maintain rigid identity consistency during complex, long-sequence actions. To address these limitations, we propose KeyID, a training-free IPVG framework that decouples the synthesis of video dynamics from the injection of identity. Specifically, KeyID comprises two components: (1) Reference-Aware Video Generation, which produces an identity-agnostic video draft aligned with multiple references, and (2) Identity-Preserved Keyframe Editing, which integrates the target identity via sparse keyframe correction and subsequent motion interpolation. By shifting from dense frame-level supervision to sparse keyframe-level refinement, KeyID effectively resolves the capacity conflict between prompt adherence and identity fidelity. Crucially, our modular design allows seamless extension to multi-subject references and complex sequential action generation without additional training. KeyID outperforms prior works and is validated by automatic and human evaluations on the official challenge benchmark, ultimately securing the runner-up position in the Track 2 (Sequential Action) of the ACM Multimedia 2026 IPVG Grand Challenge. Source code is available at https://github.com/WISLab-GDUT/KeyID.
Generating personalized dance videos from a reference image, text prompt, and audio track requires music-conditioned body motion. Singing-and-dancing adds a second requirement: the visible subject must also articulate the vocals. Existing music-conditioned methods focus primarily on choreography, while speech-driven models generally assume that the visible subject produces the input voice, leaving this combined setting largely underexplored. We introduce SingDance, a unified video diffusion framework that formulates controllable vocal articulation as a semantic role: the visible subject is either the source, who produces the vocal signal, or the listener, who receives it from an off-screen performer. Hard-compact routing selects task-relevant speech, music, and role conditions, which are composed through frame-wise joint audio injection; source and listener retain the same speech pathway. Training uses asymmetric supervision: on-screen speaking and curated off-screen conversational-response videos establish role control, while instrumental and song-based dancing-only videos establish music-conditioned body motion. The target Song/Source configuration is never observed during training. At inference, assigning the source role to a song composes separately learned articulation and song-conditioned dance capabilities, enabling compositional zero-shot singing-and-dancing. Experiments demonstrate strong motion--beat alignment and visual fidelity, reliable paired switching of vocal articulation while preserving music-aligned body motion, and highly competitive lip synchronization with substantially fewer generation-time parameters than the strongest speech-driven baseline evaluated.
展开完整摘要收起摘要↓
Generating personalized dance videos from a reference image, text prompt, and audio track requires music-conditioned body motion. Singing-and-dancing adds a second requirement: the visible subject must also articulate the vocals. Existing music-conditioned methods focus primarily on choreography, while speech-driven models generally assume that the visible subject produces the input voice, leaving this combined setting largely underexplored. We introduce SingDance, a unified video diffusion framework that formulates controllable vocal articulation as a semantic role: the visible subject is either the source, who produces the vocal signal, or the listener, who receives it from an off-screen performer. Hard-compact routing selects task-relevant speech, music, and role conditions, which are composed through frame-wise joint audio injection; source and listener retain the same speech pathway. Training uses asymmetric supervision: on-screen speaking and curated off-screen conversational-response videos establish role control, while instrumental and song-based dancing-only videos establish music-conditioned body motion. The target Song/Source configuration is never observed during training. At inference, assigning the source role to a song composes separately learned articulation and song-conditioned dance capabilities, enabling compositional zero-shot singing-and-dancing. Experiments demonstrate strong motion--beat alignment and visual fidelity, reliable paired switching of vocal articulation while preserving music-aligned body motion, and highly competitive lip synchronization with substantially fewer generation-time parameters than the strongest speech-driven baseline evaluated.
作者Feifan Zhang, Yuyang Du, Xiaoyan Liu, Soung Chang Liew
Generative image semantic communication converts an image into a text description and then performs text-to-image reconstruction at the receiver via diffusion-based generative models. This paradigm has attracted broad attention due to its extremely low bandwidth cost. However, existing methods still face two critical bottlenecks across image-to-text (I2T) semantic extraction at the transmitter and text-to-image (T2I) semantic reconstruction at the receiver: (i) semantic loss and distortion in I2T, where holistic image descriptions may omit fine-grained object attributes and spatial-position information, causing the generated text to deviate from the original image semantics; and (ii) insufficient semantic faithfulness in T2I, where even with the same semantically faithful text description, different initial noise settings may lead diffusion-based reconstruction to produce images with different levels of semantic consistency with the original image. These issues jointly limit the semantic faithfulness of image reconstruction. To address them, we propose TISC, a text-driven image semantic communication framework tailored for faithful reconstruction. TISC incorporates two key designs: (1) Tree-Structured Attribute Semantic Extraction (TSASE), which decomposes semantic extraction into global scene, background, and object-level attribute descriptions, covering spatial position, shape/pose, color, material, and other physical attributes for each detected object; and (2) an Initial Noise Optimization (INO) mechanism, which selects an initial noise seed at the transmitter according to a comprehensive similarity score that jointly considers visual and semantic consistency. Experiments on multiple datasets show that TSASE improves object-position recovery and semantic description faithfulness, while the INO parameter study supports the adopted configuration for noise selection.
展开完整摘要收起摘要↓
Generative image semantic communication converts an image into a text description and then performs text-to-image reconstruction at the receiver via diffusion-based generative models. This paradigm has attracted broad attention due to its extremely low bandwidth cost. However, existing methods still face two critical bottlenecks across image-to-text (I2T) semantic extraction at the transmitter and text-to-image (T2I) semantic reconstruction at the receiver: (i) semantic loss and distortion in I2T, where holistic image descriptions may omit fine-grained object attributes and spatial-position information, causing the generated text to deviate from the original image semantics; and (ii) insufficient semantic faithfulness in T2I, where even with the same semantically faithful text description, different initial noise settings may lead diffusion-based reconstruction to produce images with different levels of semantic consistency with the original image. These issues jointly limit the semantic faithfulness of image reconstruction. To address them, we propose TISC, a text-driven image semantic communication framework tailored for faithful reconstruction. TISC incorporates two key designs: (1) Tree-Structured Attribute Semantic Extraction (TSASE), which decomposes semantic extraction into global scene, background, and object-level attribute descriptions, covering spatial position, shape/pose, color, material, and other physical attributes for each detected object; and (2) an Initial Noise Optimization (INO) mechanism, which selects an initial noise seed at the transmitter according to a comprehensive similarity score that jointly considers visual and semantic consistency. Experiments on multiple datasets show that TSASE improves object-position recovery and semantic description faithfulness, while the INO parameter study supports the adopted configuration for noise selection.
In over-the-counter corporate bond markets, dealers compete for client trades by quoting bid and ask prices. Tighter quotes attract more business, but also informed customers more likely to trade ahead of adverse price moves, leaving the dealer holding the risk. As dealers increasingly use machine learning to set quotes, they retrain these models on the trades their own quotes attract, creating a feedback loop in which each model reshapes the market that generates its next training data. The question is therefore not only whether a quoting model performs well, but whether the market it creates stays stable as the model learns from it. Existing performative prediction theory gives a sharp stability condition, yet expresses it through abstract properties of the learning objective a trading desk cannot measure before deployment. We introduce REFLEX, a framework that replaces those unobservable quantities with three measurable features of dealer behavior: how strongly trading volume responds to tighter quotes, how sharply the dealer's objective bends around its optimum, and how quickly informed flow increases as spreads narrow. REFLEX combines these into a single retraining modulus, a pre-deployment stability margin estimated from a desk's own quote and execution history that predicts whether repeated retraining will converge or amplify itself. In simulation, predicted and measured stability agree within 8%, and competing dealers increase instability by 1.74x with two and 3.16x with three, as predicted. Where ordinary retraining becomes unstable at modulus 1.21, a structurally anchored correction converges as blind retraining collapses. Calibrated over 36 years of public market data, stability headroom falls roughly 4.4x for investment grade and 4.3x for high yield from calm to crisis regimes. Ultimately, REFLEX turns an abstract convergence theorem into a market-level safety margin.
展开完整摘要收起摘要↓
In over-the-counter corporate bond markets, dealers compete for client trades by quoting bid and ask prices. Tighter quotes attract more business, but also informed customers more likely to trade ahead of adverse price moves, leaving the dealer holding the risk. As dealers increasingly use machine learning to set quotes, they retrain these models on the trades their own quotes attract, creating a feedback loop in which each model reshapes the market that generates its next training data. The question is therefore not only whether a quoting model performs well, but whether the market it creates stays stable as the model learns from it. Existing performative prediction theory gives a sharp stability condition, yet expresses it through abstract properties of the learning objective a trading desk cannot measure before deployment. We introduce REFLEX, a framework that replaces those unobservable quantities with three measurable features of dealer behavior: how strongly trading volume responds to tighter quotes, how sharply the dealer's objective bends around its optimum, and how quickly informed flow increases as spreads narrow. REFLEX combines these into a single retraining modulus, a pre-deployment stability margin estimated from a desk's own quote and execution history that predicts whether repeated retraining will converge or amplify itself. In simulation, predicted and measured stability agree within 8%, and competing dealers increase instability by 1.74x with two and 3.16x with three, as predicted. Where ordinary retraining becomes unstable at modulus 1.21, a structurally anchored correction converges as blind retraining collapses. Calibrated over 36 years of public market data, stability headroom falls roughly 4.4x for investment grade and 4.3x for high yield from calm to crisis regimes. Ultimately, REFLEX turns an abstract convergence theorem into a market-level safety margin.
Reliable optimization is central to neural network (NN) training, yet Adam, the default optimizer for modern LLMs, rests on a fragile foundation. This thesis develops a principled grounding for Adam and motivates new designs. First, we revisit Adam's divergence--convergence debate and show the existence of a problem-dependent phase transition: with properly chosen, batch-size-dependent hyperparameters, Adam converges, whereas under small-$β_2$ regimes it can diverge. Second, we investigate why Adam substantially outperforms SGD on Transformers through Hessian structure. We find that the Hessian evolves toward a near-block-diagonal form along training, accompanied by strong block heterogeneity. We prove that this structure makes Adam's diagonal preconditioner effective. We further show that this special Hessian structure originates from consecutive multiplications of large matrix variables, and we provide a rigorous analysis based on random matrix theory. Finally, these insights motivate Adam-mini, a new optimizer that reduces Adam's memory footprint by 50% while preserving its performance. Our results also have broader implications beyond Adam: they reveal new local structures in matrix-based nonconvex problems, and also help understand and improve recent NN optimizers, such as Muon.
展开完整摘要收起摘要↓
Reliable optimization is central to neural network (NN) training, yet Adam, the default optimizer for modern LLMs, rests on a fragile foundation. This thesis develops a principled grounding for Adam and motivates new designs. First, we revisit Adam's divergence--convergence debate and show the existence of a problem-dependent phase transition: with properly chosen, batch-size-dependent hyperparameters, Adam converges, whereas under small-$β_2$ regimes it can diverge. Second, we investigate why Adam substantially outperforms SGD on Transformers through Hessian structure. We find that the Hessian evolves toward a near-block-diagonal form along training, accompanied by strong block heterogeneity. We prove that this structure makes Adam's diagonal preconditioner effective. We further show that this special Hessian structure originates from consecutive multiplications of large matrix variables, and we provide a rigorous analysis based on random matrix theory. Finally, these insights motivate Adam-mini, a new optimizer that reduces Adam's memory footprint by 50% while preserving its performance. Our results also have broader implications beyond Adam: they reveal new local structures in matrix-based nonconvex problems, and also help understand and improve recent NN optimizers, such as Muon.
作者Lea Bogensperger, Manuela Merlo, Martin Baumgartner, Michael Krauthammer, Bernard Ciraulo
Generative modeling has shown increasing promise for predicting cellular perturbation effects under chemical compound treatments. Existing approaches either model perturbation as a distribution-to-distribution mapping without explicit concentration handling, or treat concentration as a discrete class label, precluding continuous dose control. We introduce a joint flow matching approach that simultaneously models cell latents and drug concentration via a dual-timestep formulation, enabling dose-conditioned single-cell morphing through the invertibility of flow matching. The joint formulation induces a monotonic dose-response geometry in latent space and additionally supports concentration estimation from cell morphology. As proof of concept, we further demonstrate generalization to an unseen dose held out during training. Empirically, our method achieves competitive or improved per-concentration metrics on two compounds compared with representative baselines, while enabling capabilities structurally unavailable to discrete-class methods.
展开完整摘要收起摘要↓
Generative modeling has shown increasing promise for predicting cellular perturbation effects under chemical compound treatments. Existing approaches either model perturbation as a distribution-to-distribution mapping without explicit concentration handling, or treat concentration as a discrete class label, precluding continuous dose control. We introduce a joint flow matching approach that simultaneously models cell latents and drug concentration via a dual-timestep formulation, enabling dose-conditioned single-cell morphing through the invertibility of flow matching. The joint formulation induces a monotonic dose-response geometry in latent space and additionally supports concentration estimation from cell morphology. As proof of concept, we further demonstrate generalization to an unseen dose held out during training. Empirically, our method achieves competitive or improved per-concentration metrics on two compounds compared with representative baselines, while enabling capabilities structurally unavailable to discrete-class methods.
Recent advances in diffusion models and Transformer architectures have led to significant progress in text-to-video generation. However, these models often suffer from semantic errors such as missing objects, incorrect attributes, or mismatched actions. Although some semantic correction methods perform optimization before sampling or refinement after sampling, how to detect and correct semantic deviations during the video generation process remains underexplored. In this paper, we introduce a training-free, interpretable mid-generation correction framework that integrates multimodal large language model (MLLM) feedback directly into the diffusion sampling loop. Our framework achieves diffusion trajectory correction by injecting semantic evaluation signals during video synthesis, enabling the model to optimize the generated content through continuous self-reflection. We propose two key modules: a Semantic Assessment Supervisor that generates intermediate preview frames for semantic evaluations and deviation diagnostics, and a Semantic Modification Assistant that corrects semantic drift during inference via a controllable latent trajectory intervention. Our method improves semantic alignment, visual fidelity, and temporal consistency without modifying model parameters. We validate the effectiveness of our approach through extensive experiments across multiple benchmarks.
展开完整摘要收起摘要↓
Recent advances in diffusion models and Transformer architectures have led to significant progress in text-to-video generation. However, these models often suffer from semantic errors such as missing objects, incorrect attributes, or mismatched actions. Although some semantic correction methods perform optimization before sampling or refinement after sampling, how to detect and correct semantic deviations during the video generation process remains underexplored. In this paper, we introduce a training-free, interpretable mid-generation correction framework that integrates multimodal large language model (MLLM) feedback directly into the diffusion sampling loop. Our framework achieves diffusion trajectory correction by injecting semantic evaluation signals during video synthesis, enabling the model to optimize the generated content through continuous self-reflection. We propose two key modules: a Semantic Assessment Supervisor that generates intermediate preview frames for semantic evaluations and deviation diagnostics, and a Semantic Modification Assistant that corrects semantic drift during inference via a controllable latent trajectory intervention. Our method improves semantic alignment, visual fidelity, and temporal consistency without modifying model parameters. We validate the effectiveness of our approach through extensive experiments across multiple benchmarks.
作者Ye Lu, Shen Wang, Zhaoyang Zhang, Yihan Yan, Li Liu, Runze Liu, Fanghui Sun
Model Inversion Attacks (MIAs) aim to reconstruct representative training samples of target identities from face recognition models, exposing critical security vulnerabilities. Existing methods typically rely on indirect guidance or highly stochastic guidance, making it difficult to stably optimize generation trajectories toward target facial images. In this paper, we propose Steering Flow Model Inversion (SFMI), a novel two-stage white-box model inversion method that reformulates inversion as a trajectory-steering task. Specifically, Step I, Learning a Generic Flow Matching Prior, pre-trains a generic unconditional Flow Matching model to encode the manifold of human faces as a robust prior. Step II, Attacking with Progressive Guidance Scheduler (PGS), injects time-dependent target-specific gradients during sampling. By backpropagating through the target model to obtain gradients from intermediate generated states, PGS progressively injects adaptive guidance signals into the vector field. This process effectively steers the current generative flow from random noise toward the high-density regions of the target class. Under an identity-disjoint cross-evaluation setting using the CelebA dataset, SFMI achieves an ACC of 0.9248, an FID of 22.61, and an LPIPS of 0.3874 on the ArcFace target. Extensive experiments on multiple target models demonstrate that SFMI achieves competitive state-of-the-art performance in attack success and visual fidelity under the evaluated white-box protocol.
展开完整摘要收起摘要↓
Model Inversion Attacks (MIAs) aim to reconstruct representative training samples of target identities from face recognition models, exposing critical security vulnerabilities. Existing methods typically rely on indirect guidance or highly stochastic guidance, making it difficult to stably optimize generation trajectories toward target facial images. In this paper, we propose Steering Flow Model Inversion (SFMI), a novel two-stage white-box model inversion method that reformulates inversion as a trajectory-steering task. Specifically, Step I, Learning a Generic Flow Matching Prior, pre-trains a generic unconditional Flow Matching model to encode the manifold of human faces as a robust prior. Step II, Attacking with Progressive Guidance Scheduler (PGS), injects time-dependent target-specific gradients during sampling. By backpropagating through the target model to obtain gradients from intermediate generated states, PGS progressively injects adaptive guidance signals into the vector field. This process effectively steers the current generative flow from random noise toward the high-density regions of the target class. Under an identity-disjoint cross-evaluation setting using the CelebA dataset, SFMI achieves an ACC of 0.9248, an FID of 22.61, and an LPIPS of 0.3874 on the ArcFace target. Extensive experiments on multiple target models demonstrate that SFMI achieves competitive state-of-the-art performance in attack success and visual fidelity under the evaluated white-box protocol.
Outliers can substantially distort Gaussian process regression (GPR) due to its conventional Gaussian observation likelihood, leading to inaccurate model learning and prediction. To address this limitation, this article introduces a generative GPR model that captures observation-specific contamination and adaptively mitigates the influence of outliers. Subsequently, a variational generalized expectation-maximization procedure is used to learn the latent variables and GPR model parameters. Experiments on synthetic and real datasets under different contamination settings demonstrate that the proposed method remains competitive with-and in several cases outperforms-robust GPR baselines in prediction accuracy. Moreover, the proposed method shares the cubic computational scaling of the compared GPR methods.
展开完整摘要收起摘要↓
Outliers can substantially distort Gaussian process regression (GPR) due to its conventional Gaussian observation likelihood, leading to inaccurate model learning and prediction. To address this limitation, this article introduces a generative GPR model that captures observation-specific contamination and adaptively mitigates the influence of outliers. Subsequently, a variational generalized expectation-maximization procedure is used to learn the latent variables and GPR model parameters. Experiments on synthetic and real datasets under different contamination settings demonstrate that the proposed method remains competitive with-and in several cases outperforms-robust GPR baselines in prediction accuracy. Moreover, the proposed method shares the cubic computational scaling of the compared GPR methods.
Statistical inverse problems have garnered significant attention in recent years due to the growing importance of statistical learning theory and functional analytic approaches in the fields of machine learning and artificial intelligence. In this paper, we investigate the stable approximation of the element $u^{\dagger}$ that satisfies the equation $Au = g$, where $A$ is a linear operator that maps a Banach space into an appropriate function space. The function $g$ is observed only through independently and identically distributed data points that are corrupted by noise and assumed to follow an unknown distribution $ρ$. We employ the Tikhonov regularization scheme, leveraging statistical learning techniques and the framework of reproducing kernel Banach spaces to estimate the solution. We establish convergence and derive the convergence rate of the estimated solution with respect to the true solution as the number of data points increases, with the rate expressed in probabilistic terms. The theoretical findings are further supported by numerical experiments that demonstrate the effectiveness of the proposed approach.
展开完整摘要收起摘要↓
Statistical inverse problems have garnered significant attention in recent years due to the growing importance of statistical learning theory and functional analytic approaches in the fields of machine learning and artificial intelligence. In this paper, we investigate the stable approximation of the element $u^{\dagger}$ that satisfies the equation $Au = g$, where $A$ is a linear operator that maps a Banach space into an appropriate function space. The function $g$ is observed only through independently and identically distributed data points that are corrupted by noise and assumed to follow an unknown distribution $ρ$. We employ the Tikhonov regularization scheme, leveraging statistical learning techniques and the framework of reproducing kernel Banach spaces to estimate the solution. We establish convergence and derive the convergence rate of the estimated solution with respect to the true solution as the number of data points increases, with the rate expressed in probabilistic terms. The theoretical findings are further supported by numerical experiments that demonstrate the effectiveness of the proposed approach.