arXiv subjects

Sijie Zhu

Publications and source records attributed to Sijie Zhu.

At least 19 recordsLinked to original sources

arXiv2026

On the Position Bias of On-Policy Distillation

On-Policy Distillation (OPD) improves the learning efficiency of standard reinforcement learning through dense, token-level supervision from teachers. In the standard KL objective of OPD, token-level losses are uniformly averaged, implying equal weights for all tokens. However, we discover that not all tokens are created equal: as student rollouts grow longer, they deviate further from the teacher's distribution, leading to degraded supervision quality at later positions. As a result, OPD using only the first 30% of tokens can perform comparably to using all tokens, whereas OPD using only the last 30% of tokens barely learns anything. In this work, we provide a principled understanding of this issue through the lens of constrained optimization. Based on these insights, we derive Importance-Weighted On-Policy Distillation (IW-OPD), in which the weight assigned to each token depends on the accumulated discrepancy between the student's and teacher's distributions, naturally upweighting earlier tokens and downweighting later ones with larger deviations. We show that IW-OPD converges significantly faster than OPD, with better learning efficiency, and achieves better final performance than standard OPD in both same-size and cross-scale settings, improving performance up to 6.9 points on AIME-2025.

cs.LG↗

arXiv2026

VlogReward: Learning Multi-Dimensional Evaluation for Vlog Editing

The rapid rise of vlogs as a personalized storytelling medium has created a demand for automated systems to evaluate and refine vlog editing plans. However, vlog assessment is highly subjective and remains challenging due to a lack of standardized criteria, dataset and benchmark, and effective reward models. To address these challenges, we define a comprehensive vlog evaluation framework guided by professional vlog creators and product managers, establishing a taxonomy of six key dimensions, i.e., Creativity, Consistency, Concept Design, Cinematography, Narration, and Pacing. Subsequently, we curate a large-scale dataset of 100k vlog edits and a dedicated benchmark, VRMBench, to evaluate the vlog rewarding capabilities of Multimodal Large Language Models (MLLMs). Finally, we present VlogReward, a robust vlog reward model that can provide both fine-grained multi-dimensional scores and actionable feedback for iterative refinement. Technically, we enhance the Group Relative Policy Optimization (GRPO) framework by introducing an adjustable inter-group comparison reward, which mitigates the "direction blindness" issue of standard GRPO and enables the model to better distinguish varied-quality edits. VlogReward achieves state-of-the-art results that significantly outperform existing MLLMs, including GPT-5 and Gemini-3-Pro. We hope that our study can help vlog creators and foster automated vlog evaluation and refinement systems.

cs.AI↗

arXiv2026

VEBench:Benchmarking Large Multimodal Models for Real-World Video Editing

Real-world video editing demands not only expert knowledge of cinematic techniques but also multimodal reasoning to select, align, and combine footage into coherent narratives. While recent Large Multimodal Models (LMMs) have shown remarkable progress in general video understanding, their abilities in multi-video reasoning and operational editing workflows remain largely unexplored. We introduce VEBENCH, the first comprehensive benchmark designed to evaluate both editing knowledge understanding and operational reasoning in realistic video editing scenarios. VEBENCH contains 3.9K high-quality edited videos (over 257 hours) and 3,080 human-verified QA pairs, built through a three-round human-AI collaborative annotation pipeline that ensures precise temporal labeling and semantic consistency. It features two complementary QA tasks: 1) Video Editing Technique Recognition, assessing models' ability to identify 7 editing techniques using multimodal cues; and 2) Video Editing Operation Simulation, modeling real-world editing workflows by requiring the selection and temporal localization of relevant clips from multiple candidates. Extensive experiments across proprietary (e.g., Gemini-2.5-Pro) and open-source LMMs reveal a large gap between current model performance and human-level editing cognition. These results highlight the urgent need for bridging video understanding with creative operational reasoning. We envision VEBENCH as a foundation for advancing intelligent video editing systems and driving future research on complex reasoning.

cs.CV↗

arXiv2026

Vidi2.5: Large Multimodal Models for Video Understanding and Creation

Video has emerged as the primary medium for communication and creativity on the Internet, driving strong demand for scalable, high-quality video production. Vidi models continue to evolve toward next-generation video creation and have achieved state-of-the-art performance in multimodal temporal retrieval (TR). In its second release, Vidi2 advances video understanding with fine-grained spatio-temporal grounding (STG) and extends its capability to video question answering (Video QA), enabling comprehensive multimodal reasoning. Given a text query, Vidi2 can identify not only the corresponding timestamps but also the bounding boxes of target objects within the output time ranges. To enable comprehensive evaluation of STG, we introduce a new benchmark, VUE-STG, which offers critical improvements over existing STG datasets. In addition, we upgrade the previous VUE-TR benchmark to VUE-TR-V2, achieving a more balanced duration and query distribution. Remarkably, the Vidi2 model substantially outperforms leading proprietary systems, such as Gemini 3 Pro Preview and GPT-5, on both VUE-TR-V2 and VUE-STG, while achieving competitive results with popular open-source models with similar scale on video QA benchmarks. The latest Vidi2.5 offers significantly stronger STG capability and slightly better TR and Video QA performance over Vidi2. This update also introduces a Vidi2.5-Think model to handle plot understanding with complex plot reasoning. To comprehensively evaluate the performance of plot understanding, we propose VUE-PLOT benchmark with two tracks, Character and Reasoning. Notably, Vidi2.5-Think outperforms Gemini 3 Pro Preview on fine-grained character understanding with comparable performance on complex plot reasoning. Furthermore, we demonstrate the effectiveness of Vidi2.5 on a challenging real-world application, video editing planning.

cs.CV↗

arXiv2025

Thinking With Bounding Boxes: Enhancing Spatio-Temporal Video Grounding via Reinforcement Fine-Tuning

Spatio-temporal video grounding (STVG) requires localizing a target object in untrimmed videos both temporally and spatially from natural language descriptions. Despite their strong language understanding, multimodal large language models (MLLMs) underperform on STVG due to misaligned training objectives and weak fine-grained region-word alignment in standard visual encoders. To address this, we propose STVG-o1, the first framework that enables off-the-shelf MLLMs to achieve state-of-the-art STVG performance without any architectural modifications. Our method introduces a bounding-box chain-of-thought mechanism that explicitly reasons about spatio-temporal locations in an intermediate step before producing the final prediction. We further design a multi-dimensional reinforcement reward function consisting of format, consistency, temporal, spatial, and think rewards, which provides geometry-aware supervision through reinforcement fine-tuning. Evaluated on HCSTVG-v1/v2 and VidSTG, STVG-o1 sets new state-of-the-art results on HCSTVG, outperforming the best task-specific method by 7.3\% m\_tIoU on HCSTVG-v1, matching specialized models on VidSTG, and surpassing all existing MLLM-based approaches by large margins. It also demonstrates strong open-vocabulary generalization across datasets, establishing MLLMs as viable and powerful backbones for precise spatio-temporal grounding. Our code and models will be released.

cs.CV↗

arXiv2025

D-Attn: Decomposed Attention for Large Vision-and-Language Models

Large vision-and-language models (LVLMs) have traditionally integrated visual and textual tokens by concatenating them into a single homogeneous input for large language models (LLMs), thereby maximally preserving the pre-trained language capabilities. However, this constrained architecture for visual and textual tokens restricts the design space for processing visual tokens, potentially leading to suboptimal performance and efficiency. In this paper, we propose Decomposed Attention (D-Attn), a more flexible attention architecture for LVLMs, which enables modification of visual token operations without affecting textual-to-textual attention. D-Attn decomposes the 1-D causal self-attention of LVLMs into visual-to-visual, textual-to-visual, and textual-to-textual attentions, and the visual and textual output tokens from the decomposed attentions are merged with a carefully derived weighting strategy, namely $α$-weighting. Taking advantage of the flexibility, we are able to introduce two critical improvements in visual token processing while maintaining the capacity of pre-trained LLMs: 1) We rectify the biased positional encoding in textual-to-visual attention to boost visual understanding performance. 2) We diagonalize visual-to-visual attention to reduce computation complexity from $O(|V|^2)$ to $O(|V|)$ for $|V|$ visual tokens without compromising performance. Extensive experiments and analysis validate the effectiveness of D-Attn, demonstrating significant improvements on multiple image benchmarks while significantly reducing computational costs (\eg, $5\times$ faster). Code will be available at https://github.com/bytedance/DecomposedAttention.

cs.CV↗

arXiv2025

Vidi: Large Multimodal Models for Video Understanding and Editing

Humans naturally share information with those they are connected to, and video has become one of the dominant mediums for communication and expression on the Internet. To support the creation of high-quality large-scale video content, a modern pipeline requires a comprehensive understanding of both the raw input materials (e.g., the unedited footage captured by cameras) and the editing components (e.g., visual effects). In video editing scenarios, models must process multiple modalities (e.g., vision, audio, text) with strong background knowledge and handle flexible input lengths (e.g., hour-long raw videos), which poses significant challenges for traditional models. In this report, we introduce Vidi, a family of Large Multimodal Models (LMMs) for a wide range of video understand editing scenarios. The first release focuses on temporal retrieval, i.e., identifying the time ranges within the input videos corresponding to a given text query, which plays a critical role in intelligent editing. The model is capable of processing hour-long videos with strong temporal understanding capability, e.g., retrieve time ranges for certain queries. To support a comprehensive evaluation in real-world scenarios, we also present the VUE-TR benchmark, which introduces five key advancements. 1) Video duration: significantly longer than videos of existing temporal retrival datasets, 2) Audio support: includes audio-based queries, 3) Query format: diverse query lengths/formats, 4) Annotation quality: ground-truth time ranges are manually annotated. 5) Evaluation metric: a refined IoU metric to support evaluation over multiple time ranges. Remarkably, Vidi significantly outperforms leading proprietary models, e.g., GPT-4o and Gemini, on the temporal retrieval task, indicating its superiority in video editing scenarios.

cs.CV↗

arXiv2025

SuperEdit: Rectifying and Facilitating Supervision for Instruction-Based Image Editing

Due to the challenges of manually collecting accurate editing data, existing datasets are typically constructed using various automated methods, leading to noisy supervision signals caused by the mismatch between editing instructions and original-edited image pairs. Recent efforts attempt to improve editing models through generating higher-quality edited images, pre-training on recognition tasks, or introducing vision-language models (VLMs) but fail to resolve this fundamental issue. In this paper, we offer a novel solution by constructing more effective editing instructions for given image pairs. This includes rectifying the editing instructions to better align with the original-edited image pairs and using contrastive editing instructions to further enhance their effectiveness. Specifically, we find that editing models exhibit specific generation attributes at different inference steps, independent of the text. Based on these prior attributes, we define a unified guide for VLMs to rectify editing instructions. However, there are some challenging editing scenarios that cannot be resolved solely with rectified instructions. To this end, we further construct contrastive supervision signals with positive and negative instructions and introduce them into the model training using triplet loss, thereby further facilitating supervision effectiveness. Our method does not require the VLM modules or pre-training tasks used in previous work, offering a more direct and efficient way to provide better supervision signals, and providing a novel, simple, and effective solution for instruction-based image editing. Results on multiple benchmarks demonstrate that our method significantly outperforms existing approaches. Compared with previous SOTA SmartEdit, we achieve 9.19% improvements on the Real-Edit benchmark with 30x less training data and 13x smaller model size.

cs.CV↗

arXiv2025

Multi-Reward as Condition for Instruction-based Image Editing

High-quality training triplets (instruction, original image, edited image) are essential for instruction-based image editing. Predominant training datasets (e.g., InsPix2Pix) are created using text-to-image generative models (e.g., Stable Diffusion, DALL-E) which are not trained for image editing. Accordingly, these datasets suffer from inaccurate instruction following, poor detail preserving, and generation artifacts. In this paper, we propose to address the training data quality issue with multi-perspective reward data instead of refining the ground-truth image quality. 1) we first design a quantitative metric system based on best-in-class LVLM (Large Vision Language Model), i.e., GPT-4o in our case, to evaluate the generation quality from 3 perspectives, namely, instruction following, detail preserving, and generation quality. For each perspective, we collected quantitative score in $0\sim 5$ and text descriptive feedback on the specific failure points in ground-truth edited images, resulting in a high-quality editing reward dataset, i.e., RewardEdit20K. 2) We further proposed a novel training framework to seamlessly integrate the metric output, regarded as multi-reward, into editing models to learn from the imperfect training triplets. During training, the reward scores and text descriptions are encoded as embeddings and fed into both the latent space and the U-Net of the editing models as auxiliary conditions. 3) We also build a challenging evaluation benchmark with real-world images/photos and diverse editing instructions, named Real-Edit. Experiments indicate that our multi-reward conditioned model outperforms its no-reward counterpart on two popular editing pipelines, i.e., InsPix2Pix and SmartEdit. Code is released at https://github.com/bytedance/Multi-Reward-Editing.

cs.CV↗

arXiv2025

Where do Large Vision-Language Models Look at when Answering Questions?

Large Vision-Language Models (LVLMs) have shown promising performance in vision-language understanding and reasoning tasks. However, their visual understanding behaviors remain underexplored. A fundamental question arises: to what extent do LVLMs rely on visual input, and which image regions contribute to their responses? It is non-trivial to interpret the free-form generation of LVLMs due to their complicated visual architecture (e.g., multiple encoders and multi-resolution) and variable-length outputs. In this paper, we extend existing heatmap visualization methods (e.g., iGOS++) to support LVLMs for open-ended visual question answering. We propose a method to select visually relevant tokens that reflect the relevance between generated answers and input image. Furthermore, we conduct a comprehensive analysis of state-of-the-art LVLMs on benchmarks designed to require visual information to answer. Our findings offer several insights into LVLM behavior, including the relationship between focus region and answer correctness, differences in visual attention across architectures, and the impact of LLM scale on visual understanding. The code and data are available at https://github.com/bytedance/LVLM_Interpretation.

cs.CV↗

arXiv2025

Edit3K: Universal Representation Learning for Video Editing Components

This paper focuses on understanding the predominant video creation pipeline, i.e., compositional video editing with six main types of editing components, including video effects, animation, transition, filter, sticker, and text. In contrast to existing visual representation learning of visual materials (i.e., images/videos), we aim to learn visual representations of editing actions/components that are generally applied on raw materials. We start by proposing the first large-scale dataset for editing components of video creation, which covers about $3,094$ editing components with $618,800$ videos. Each video in our dataset is rendered by various image/video materials with a single editing component, which supports atomic visual understanding of different editing components. It can also benefit several downstream tasks, e.g., editing component recommendation, editing component recognition/retrieval, etc. Existing visual representation methods perform poorly because it is difficult to disentangle the visual appearance of editing components from raw materials. To that end, we benchmark popular alternative solutions and propose a novel method that learns to attend to the appearance of editing components regardless of raw materials. Our method achieves favorable results on editing component retrieval/recognition compared to the alternative solutions. A user study is also conducted to show that our representations cluster visually similar editing components better than other alternatives. Furthermore, our learned representations used to transition recommendation tasks achieve state-of-the-art results on the AutoTransition dataset. The code and dataset are available at https://github.com/GX77/Edit3K .

cs.CV↗

arXiv2024

Beyond Raw Videos: Understanding Edited Videos with Large Multimodal Model

The emerging video LMMs (Large Multimodal Models) have achieved significant improvements on generic video understanding in the form of VQA (Visual Question Answering), where the raw videos are captured by cameras. However, a large portion of videos in real-world applications are edited videos, \textit{e.g.}, users usually cut and add effects/modifications to the raw video before publishing it on social media platforms. The edited videos usually have high view counts but they are not covered in existing benchmarks of video LMMs, \textit{i.e.}, ActivityNet-QA, or VideoChatGPT benchmark. In this paper, we leverage the edited videos on a popular short video platform, \textit{i.e.}, TikTok, and build a video VQA benchmark (named EditVid-QA) covering four typical editing categories, i.e., effect, funny, meme, and game. Funny and meme videos benchmark nuanced understanding and high-level reasoning, while effect and game evaluate the understanding capability of artificial design. Most of the open-source video LMMs perform poorly on the EditVid-QA benchmark, indicating a huge domain gap between edited short videos on social media and regular raw videos. To improve the generalization ability of LMMs, we collect a training set for the proposed benchmark based on both Panda-70M/WebVid raw videos and small-scale TikTok/CapCut edited videos, which boosts the performance on the proposed EditVid-QA benchmark, indicating the effectiveness of high-quality training data. We also identified a serious issue in the existing evaluation protocol using the GPT-3.5 judge, namely a "sorry" attack, where a sorry-style naive answer can achieve an extremely high rating from the GPT judge, e.g., over 4.3 for correctness score on VideoChatGPT evaluation protocol. To avoid the "sorry" attacks, we evaluate results with GPT-4 judge and keyword filtering. The dataset is released at https://github.com/XenonLamb/EditVid-QA.

cs.CV↗

arXiv2024

CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts

Recent advancements in Multimodal Large Language Models (LLMs) have focused primarily on scaling by increasing text-image pair data and enhancing LLMs to improve performance on multimodal tasks. However, these scaling approaches are computationally expensive and overlook the significance of improving model capabilities from the vision side. Inspired by the successful applications of Mixture-of-Experts (MoE) in LLMs, which improves model scalability during training while keeping inference costs similar to those of smaller models, we propose CuMo. CuMo incorporates Co-upcycled Top-K sparsely-gated Mixture-of-experts blocks into both the vision encoder and the MLP connector, thereby enhancing the multimodal LLMs with minimal additional activated parameters during inference. CuMo first pre-trains the MLP blocks and then initializes each expert in the MoE block from the pre-trained MLP block during the visual instruction tuning stage. Auxiliary losses are used to ensure a balanced loading of experts. CuMo outperforms state-of-the-art multimodal LLMs across various VQA and visual-instruction-following benchmarks using models within each model size group, all while training exclusively on open-sourced datasets. The code and model weights for CuMo are open-sourced at https://github.com/SHI-Labs/CuMo.

cs.CV↗

arXiv2023

Deep Learning-Based Human Pose Estimation: A Survey

Human pose estimation aims to locate the human body parts and build human body representation (e.g., body skeleton) from input data such as images and videos. It has drawn increasing attention during the past decade and has been utilized in a wide range of applications including human-computer interaction, motion analysis, augmented reality, and virtual reality. Although the recently developed deep learning-based solutions have achieved high performance in human pose estimation, there still remain challenges due to insufficient training data, depth ambiguities, and occlusion. The goal of this survey paper is to provide a comprehensive review of recent deep learning-based solutions for both 2D and 3D pose estimation via a systematic analysis and comparison of these solutions based on their input data and inference procedures. More than 250 research papers since 2014 are covered in this survey. Furthermore, 2D and 3D human pose estimation datasets and evaluation metrics are included. Quantitative performance comparisons of the reviewed methods on popular datasets are summarized and discussed. Finally, the challenges involved, applications, and future research directions are concluded. A regularly updated project page is provided: \url{https://github.com/zczcwh/DL-HPE}

cs.CV↗

arXiv2023

TopNet: Transformer-based Object Placement Network for Image Compositing

We investigate the problem of automatically placing an object into a background image for image compositing. Given a background image and a segmented object, the goal is to train a model to predict plausible placements (location and scale) of the object for compositing. The quality of the composite image highly depends on the predicted location/scale. Existing works either generate candidate bounding boxes or apply sliding-window search using global representations from background and object images, which fail to model local information in background images. However, local clues in background images are important to determine the compatibility of placing the objects with certain locations/scales. In this paper, we propose to learn the correlation between object features and all local background features with a transformer module so that detailed information can be provided on all possible location/scale configurations. A sparse contrastive loss is further proposed to train our model with sparse supervision. Our new formulation generates a 3D heatmap indicating the plausibility of all location/scale combinations in one network forward pass, which is over 10 times faster than the previous sliding-window method. It also supports interactive search when users provide a pre-defined location or scale. The proposed method can be trained with explicit annotation or in a self-supervised manner using an off-the-shelf inpainting model, and it outperforms state-of-the-art methods significantly. The user study shows that the trained model generalizes well to real-world images with diverse challenging scenes and object categories.

cs.CV↗

arXiv2023

$R^{2}$Former: Unified $R$etrieval and $R$eranking Transformer for Place Recognition

Visual Place Recognition (VPR) estimates the location of query images by matching them with images in a reference database. Conventional methods generally adopt aggregated CNN features for global retrieval and RANSAC-based geometric verification for reranking. However, RANSAC only employs geometric information but ignores other possible information that could be useful for reranking, e.g. local feature correlations, and attention values. In this paper, we propose a unified place recognition framework that handles both retrieval and reranking with a novel transformer model, named $R^{2}$Former. The proposed reranking module takes feature correlation, attention value, and xy coordinates into account, and learns to determine whether the image pair is from the same location. The whole pipeline is end-to-end trainable and the reranking module alone can also be adopted on other CNN or transformer backbones as a generic component. Remarkably, $R^{2}$Former significantly outperforms state-of-the-art methods on major VPR datasets with much less inference time and memory consumption. It also achieves the state-of-the-art on the hold-out MSLS challenge set and could serve as a simple yet strong solution for real-world large-scale applications. Experiments also show vision transformer tokens are comparable and sometimes better than CNN local features on local matching. The code is released at https://github.com/Jeff-Zilence/R2Former.

cs.CV↗

arXiv2022

Consistency-based Active Learning for Object Detection

Active learning aims to improve the performance of task model by selecting the most informative samples with a limited budget. Unlike most recent works that focused on applying active learning for image classification, we propose an effective Consistency-based Active Learning method for object Detection (CALD), which fully explores the consistency between original and augmented data. CALD has three appealing benefits. (i) CALD is systematically designed by investigating the weaknesses of existing active learning methods, which do not take the unique challenges of object detection into account. (ii) CALD unifies box regression and classification with a single metric, which is not concerned by active learning methods for classification. CALD also focuses on the most informative local region rather than the whole image, which is beneficial for object detection. (iii) CALD not only gauges individual information for sample selection, but also leverages mutual information to encourage a balanced data distribution. Extensive experiments show that CALD significantly outperforms existing state-of-the-art task-agnostic and detection-specific active learning methods on general object detection datasets. Based on the Faster R-CNN detector, CALD consistently surpasses the baseline method (random selection) by 2.9/2.8/0.8 mAP on average on PASCAL VOC 2007, PASCAL VOC 2012, and MS COCO. Code is available at \url{https://github.com/we1pingyu/CALD}

cs.CV↗

arXiv2022

TransGeo: Transformer Is All You Need for Cross-view Image Geo-localization

The dominant CNN-based methods for cross-view image geo-localization rely on polar transform and fail to model global correlation. We propose a pure transformer-based approach (TransGeo) to address these limitations from a different perspective. TransGeo takes full advantage of the strengths of transformer related to global information modeling and explicit position information encoding. We further leverage the flexibility of transformer input and propose an attention-guided non-uniform cropping method, so that uninformative image patches are removed with negligible drop on performance to reduce computation cost. The saved computation can be reallocated to increase resolution only for informative patches, resulting in performance improvement with no additional computation cost. This "attend and zoom-in" strategy is highly similar to human behavior when observing images. Remarkably, TransGeo achieves state-of-the-art results on both urban and rural datasets, with significantly less computation cost than CNN-based methods. It does not rely on polar transform and infers faster than CNN-based methods. Code is available at https://github.com/Jeff-Zilence/TransGeo2022.

cs.CV↗