arXiv subjects

Jiayi Song

Publications and source records attributed to Jiayi Song.

18 recordsLinked to original sources

arXiv2026

EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing

High-resolution image editing is increasingly demanded in professional workflows, yet existing diffusion-based models remain constrained to resolutions below 1K due to quadratic attention complexity and prohibitive memory requirements. A prevalent workaround employs a two-stage pipeline: editing at low resolution followed by independent super-resolution. However, this approach suffers from two critical issues: information divergence, where hallucinated details contradict the original high-resolution (HR) source, and texture degradation, manifesting as over-smoothed or over-sharpened artifacts. We propose EditBridge, a diffusion bridge framework for efficient ultra high-resolution editing. Unlike conventional diffusion that regenerates from noise, we formulate refinement as structured data-to-data translation from the low-resolution (LR) edited result to its HR counterpart, explicitly conditioned on the original HR source to preserve authentic details. To efficiently incorporate HR source guidance, we introduce a prior-guided block-wise sparse attention mechanism that exploits semantic correspondence from first-stage editing to constrain cross-image interactions to spatially aligned regions, significantly reducing computational overhead. Extensive experiments demonstrate that EditBridge achieves high-fidelity editing with superior perceptual quality at resolutions up to 4K, delivering 3.6--8.4$\times$ speedup at 2K and enabling practical 4K editing in 61 seconds.

cs.CV↗

arXiv2026

LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time

Pose-driven human animation synthesizes a video of a target person from a single reference image and a driving pose stream. Real-time generation is essential for interactive applications such as live streaming, telepresence, and virtual avatars, yet diffusion-based systems require minutes to hours per clip, precluding responsive interaction. We present LiveAnimate, to our knowledge the first animation system to combine real-time streaming with stable long-form generation at billion scale, built on a 14B-parameter video Diffusion Transformer (DiT). A two-stage training pipeline first adapts a pretrained bidirectional DiT into a block-causal autoregressive generator through Reference-Anchored Teacher-Forcing Adaptation, and then reduces the sampling budget to three steps through Block-wise Self-Forcing Distillation. To preserve appearance over extended streams, we introduce Pose-Retrieval Sink Attention (PR-Sink), a bounded KV-cache mechanism combining a Static Sink that permanently anchors the first generated block, a Dynamic Sink that holds a pose-retrieved historical block, and a three-slot Rolling Window. When a pose recurs, PR-Sink restores the relevant appearance context without retaining the entire sequence, so memory and per-block latency remain constant regardless of stream duration. Together with Ulysses sequence parallelism and operator fusion, these designs enable 19.63\,FPS streaming inference on two NVIDIA H100 GPUs. On a three-minute benchmark, LiveAnimate maintains nearly constant perceptual quality and identity from the first 30 seconds to the final minute, while prior systems degrade substantially or require hours of offline computation for the same rollout. These results establish a new operating point in quality, latency, and duration for interactive full-body animation.

cs.CV↗

arXiv2026

UniSwap: Streaming Audio-Visual Identity Swapping for Talking Videos

Talking-video character replacement requires coordinated transfer of appearance and voice while preserving the source motion, scene, linguistic content, and audio-video timing. Existing methods use separately optimized models for the two modalities, making audio-visual consistency difficult to enforce. We present UniSwap, the first framework for streaming joint audio-visual identity replacement in talking videos. Given a source video, a reference image, and a reference voice clip, UniSwap transfers the reference appearance and vocal timbre within a single audio-visual diffusion transformer while preserving the source content and dynamics. To address the scarcity of aligned cross-identity training pairs, we introduce a swap-and-reconstruct pipeline that removes visual and vocal identity from real clips and uses the original clips as reconstruction targets. Starting from a bidirectional backbone, we progressively adapt the model through In-context Pretraining for joint replacement, Conditional Streaming Adaptation for block-causal KV-cached generation, and Efficient Self-forcing DMD for mitigating exposure bias and reducing sampling from 30 to 3 denoising steps per block. Efficient Multi-LoRA Switching enables the three DMD roles to share a single frozen backbone. Feature-RoPE Decomposition keeps cached positions within the training range, supporting stable long-form inference. Experiments demonstrate strong audio-visual synchronization, competitive identity preservation, efficient streaming, and stable long-form generation.

cs.CV↗

arXiv2026

ED-DiT: Physics-Guided Diffusion Pretraining for Transferable Molecular Representations from Electron Density

Pretraining has shown strong potential for learning transferable representations, yet it remains underexplored for electron-density-based molecular learning. Electron density provides a continuous three-dimensional description of molecular electronic structure, capturing both local spatial patterns and global physical quantities. This raises a key question: can electron-density fields be used for self-supervised pretraining to learn a shared representation that transfers across diverse electronic-structure-related tasks? We propose ED-DiT, a physics-guided Diffusion Transformer for self-supervised pretraining on electron-density point clouds. ED-DiT learns reusable representations by reconstructing corrupted and partially masked log-density fields across diffusion noise levels. An electron-number consistency constraint is further introduced to preserve the total electronic mass. The pretrained encoder can be adapted to property prediction, open-/closed-shell classification, molecule-electron-density retrieval, and molecule-conditioned electron-density prediction. Experiments on six EDBench tasks show that ED-DiT consistently outperforms the same architecture trained from scratch, especially under limited supervision. For molecule-conditioned electron-density prediction, it reduces RMSE from 2.2474 to 1.3753 and surpasses the available baseline. With only 10% labels, it improves orbital energy prediction RMSE from 0.0293 to 0.0138. These results demonstrate the effectiveness of physics-guided electron-density pretraining for learning transferable molecular representations.

cs.LG↗

arXiv2026

Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle

As foundation models advance and agent scaffolding becomes increasingly sophisticated, agents have demonstrated remarkable proficiency in complex, long-horizon coding tasks and even autonomous experiment execution. Despite their evolution from research assistants into autonomous research agents, these systems still exhibit significant limitations in field sensitivity, research ethics, and nuanced scientific judgment. Consequently, frontier agents remain unable to fully replace human researchers. To bridge this gap, we conceptualize the AARR (Act As a Real Researcher) benchmark series. Unlike existing benchmarks that primarily assess macro-level execution capabilities, AARR focuses on whether agents can emulate the professionalism, thoroughness, and nuanced reasoning that characterize human researchers in granular research scenarios. In this work, we propose AARRI-Bench (Act As a Real Research Intern), the first benchmark in this series. We conduct extensive experiments across frontier models and agentic systems, revealing that even the best-performing configuration (Mini-SWE-Agent with Claude Opus 4.7) achieves only 68.3\% success rate, frequently overlooking subtle yet critical details that are obvious to real human researchers. Our results indicate that developing researcher-like AI requires further exploration of research behavior, rather than merely complex scaffolding. Our data is released at https://github.com/AARR-bench/AARRI-bench.

cs.AI↗

arXiv2026

CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation

Customized image editing aims to equip pre-trained diffusion models with specific visual effects using limited paired data, typically via Low-Rank Adaptation (LoRA). As the number of desired effects grows, storing and dynamically loading numerous these effect LoRAs significantly increases deployment overhead. Furthermore, current pipelines typically cascade these effect LoRAs with acceleration modules for fast generation, which triggers severe parameter interference and results in concept bleeding and style degradation. We propose CollectionLoRA, a multi-teacher on-policy distillation framework capable of distilling the concepts of up to 50 different effect LoRAs along with few-step generation capabilities into a single LoRA. This fundamentally resolves the feature interference issue and significantly reduces deployment costs. Specifically, the method introduces (i) a Probabilistic Dual-Stream Routing mechanism that enables the model to randomly switch between data sources during training, effectively enhancing its generalization in unseen scenarios; (ii) an Asymmetric Orthogonal Prompting strategy to achieve concept isolation within the prompt space; (iii) a Coarse-to-Fine Distillation Objective to mitigate the distribution gap between the teacher and student models. Extensive evaluations show that CollectionLoRA distills all customized effects and few-step generation into a single LoRA, reducing deployment overhead while achieving concept fidelity comparable to or better than independently trained teacher models. Code: https://github.com/Qwen-Applications/CollectionLoRA

cs.CV↗

arXiv2025

FreeSwim: Revisiting Sliding-Window Attention Mechanisms for Training-Free Ultra-High-Resolution Video Generation

The quadratic time and memory complexity of the attention mechanism in modern Transformer based video generators makes end-to-end training for ultra high resolution videos prohibitively expensive. Motivated by this limitation, we introduce a training-free approach that leverages video Diffusion Transformers pretrained at their native scale to synthesize higher resolution videos without any additional training or adaptation. At the core of our method lies an inward sliding window attention mechanism, which originates from a key observation: maintaining each query token's training scale receptive field is crucial for preserving visual fidelity and detail. However, naive local window attention, unfortunately, often leads to repetitive content and exhibits a lack of global coherence in the generated results. To overcome this challenge, we devise a dual-path pipeline that backs up window attention with a novel cross-attention override strategy, enabling the semantic content produced by local attention to be guided by another branch with a full receptive field and, therefore, ensuring holistic consistency. Furthermore, to improve efficiency, we incorporate a cross-attention caching strategy for this branch to avoid the frequent computation of full 3D attention. Extensive experiments demonstrate that our method delivers ultra-high-resolution videos with fine-grained visual details and high efficiency in a training-free paradigm. Meanwhile, it achieves superior performance on VBench, even compared to training-based alternatives, with competitive or improved efficiency. Codes are available at: https://github.com/WillWu111/FreeSwim

cs.CV↗

arXiv2025

ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks

Ultra-high-resolution (UHR) remote sensing (RS) images offer rich fine-grained information but also present challenges in effective processing. Existing dynamic resolution and token pruning methods are constrained by a passive perception paradigm, suffering from increased redundancy when obtaining finer visual inputs. In this work, we explore a new active perception paradigm that enables models to revisit information-rich regions. First, we present LRS-GRO, a large-scale benchmark dataset tailored for active perception in UHR RS processing, encompassing 17 question types across global, region, and object levels, annotated via a semi-automatic pipeline. Building on LRS-GRO, we propose ZoomEarth, an adaptive cropping-zooming framework with a novel Region-Guided reward that provides fine-grained guidance. Trained via supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), ZoomEarth achieves state-of-the-art performance on LRS-GRO and, in the zero-shot setting, on three public UHR remote sensing benchmarks. Furthermore, ZoomEarth can be seamlessly integrated with downstream models for tasks such as cloud removal, denoising, segmentation, and image editing through simple tool interfaces, demonstrating strong versatility and extensibility.

cs.CV↗

arXiv2025

DeepThink3D: Enhancing Large Language Models with Programmatic Reasoning in Complex 3D Situated Reasoning Tasks

This work enhances the ability of large language models (LLMs) to perform complex reasoning in 3D scenes. Recent work has addressed the 3D situated reasoning task by invoking tool usage through large language models. Large language models call tools via APIs and integrate the generated programs through a chain of thought to solve problems based on the program results. However, due to the simplicity of the questions in the dataset, the generated program reasoning chains are relatively short. To solve this main challenge, in this paper, we introduce DeepThink3D to enhance the tool usage of LLMs in complex 3D situated reasoning tasks. Our work proposes a combinatorial and iterative evolutionary approach on the SQA3D benchmark to generate more complex questions. Building on this foundation, we fine-tune the large language model to make it more proficient in using 3D tools. By employing Direct Preference Optimization (DPO), we directly optimize the toolchain strategies generated by models, thereby enhancing their accuracy in complex tasks.

cs.AI↗

arXiv2025

A Multi-View High-Resolution Foot-Ankle Complex Point Cloud Dataset During Gait for Occlusion-Robust 3D Completion

The kinematics analysis of foot-ankle complex during gait is essential for advancing biomechanical research and clinical assessment. Collecting accurate surface geometry data from the foot and ankle during dynamic gait conditions is inherently challenging due to swing foot occlusions and viewing limitations. Thus, this paper introduces FootGait3D, a novel multi-view dataset of high-resolution ankle-foot surface point clouds captured during natural gait. Different from existing gait datasets that typically target whole-body or lower-limb motion, FootGait3D focuses specifically on the detailed modeling of the ankle-foot region, offering a finer granularity of motion data. To address this, FootGait3D consists of 8,403 point cloud frames collected from 46 subjects using a custom five-camera depth sensing system. Each frame includes a complete 5-view reconstruction of the foot and ankle (serving as ground truth) along with partial point clouds obtained from only four, three, or two views. This structured variation enables rigorous evaluation of 3D point cloud completion methods under varying occlusion levels and viewpoints. Our dataset is designed for shape completion tasks, facilitating the benchmarking of state-of-the-art single-modal (e.g., PointTr, SnowflakeNet, Anchorformer) and multi-modal (e.g., SVDFormer, PointSea, CSDN) completion networks on the challenge of recovering the full foot geometry from occluded inputs. FootGait3D has significant potential to advance research in biomechanics and multi-segment foot modeling, offering a valuable testbed for clinical gait analysis, prosthetic design, and robotics applications requiring detailed 3D models of the foot during motion. The dataset is now available at https://huggingface.co/datasets/ljw285/FootGait3D.

cs.CV↗

arXiv2025

Reflections Unlock: Geometry-Aware Reflection Disentanglement in 3D Gaussian Splatting for Photorealistic Scenes Rendering

Accurately rendering scenes with reflective surfaces remains a significant challenge in novel view synthesis, as existing methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) often misinterpret reflections as physical geometry, resulting in degraded reconstructions. Previous methods rely on incomplete and non-generalizable geometric constraints, leading to misalignment between the positions of Gaussian splats and the actual scene geometry. When dealing with real-world scenes containing complex geometry, the accumulation of Gaussians further exacerbates surface artifacts and results in blurred reconstructions. To address these limitations, in this work, we propose Ref-Unlock, a novel geometry-aware reflection modeling framework based on 3D Gaussian Splatting, which explicitly disentangles transmitted and reflected components to better capture complex reflections and enhance geometric consistency in real-world scenes. Our approach employs a dual-branch representation with high-order spherical harmonics to capture high-frequency reflective details, alongside a reflection removal module providing pseudo reflection-free supervision to guide clean decomposition. Additionally, we incorporate pseudo-depth maps and a geometry-aware bilateral smoothness constraint to enhance 3D geometric consistency and stability in decomposition. Extensive experiments demonstrate that Ref-Unlock significantly outperforms classical GS-based reflection methods and achieves competitive results with NeRF-based models, while enabling flexible vision foundation models (VFMs) driven reflection editing. Our method thus offers an efficient and generalizable solution for realistic rendering of reflective scenes. Our code is available at https://ref-unlock.github.io/.

cs.CV↗

arXiv2025

RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation

Semantic segmentation in remote sensing images is crucial for various applications, yet its performance is heavily reliant on large-scale, high-quality pixel-wise annotations, which are notoriously expensive and time-consuming to acquire. Semi-supervised semantic segmentation (SSS) offers a promising alternative to mitigate this data dependency. However, existing SSS methods often struggle with the inherent distribution mismatch between limited labeled data and abundant unlabeled data, leading to suboptimal generalization. To alleviate this issue, we attempt to introduce the Vision Foundation Models (VFMs) pre-trained on vast and diverse datasets into the SSS task since VFMs possess robust generalization capabilities that can effectively bridge this distribution gap and provide strong semantic priors for SSS. Inspired by this, we introduce RS-MTDF (Multi-Teacher Distillation and Fusion), a novel framework that leverages the powerful semantic knowledge embedded in VFMs to guide semi-supervised learning in remote sensing. Specifically, RS-MTDF employs multiple frozen VFMs (e.g., DINOv2 and CLIP) as expert teachers, utilizing feature-level distillation to align student features with their robust representations. To further enhance discriminative power, the distilled knowledge is seamlessly fused into the student decoder. Extensive experiments on three challenging remote sensing datasets demonstrate that RS-MTDF consistently achieves state-of-the-art performance. Notably, our method outperforms existing approaches across various label ratios on LoveDA and secures the highest IoU in the majority of semantic categories. These results underscore the efficacy of multi-teacher VFM guidance in significantly enhancing both generalization and semantic understanding for remote sensing segmentation. Ablation studies further validate the contribution of each proposed module.

cs.CV↗

arXiv2025

Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning

Multimodal large language models (MLLMs) have achieved strong performance on vision-language tasks but still struggle with fine-grained visual differences, leading to hallucinations or missed semantic shifts. We attribute this to limitations in both training data and learning objectives. To address these issues, we propose a controlled data generation pipeline that produces minimally edited image pairs with semantically aligned captions. Using this pipeline, we construct the Micro Edit Dataset (MED), containing over 50K image-text pairs spanning 11 fine-grained edit categories, including attribute, count, position, and object presence changes. Building on MED, we introduce a supervised fine-tuning (SFT) framework with a feature-level consistency loss that promotes stable visual embeddings under small edits. We evaluate our approach on the Micro Edit Detection benchmark, which includes carefully balanced evaluation pairs designed to test sensitivity to subtle visual variations across the same edit categories. Our method improves difference detection accuracy and reduces hallucinations compared to strong baselines, including GPT-4o. Moreover, it yields consistent gains on standard vision-language tasks such as image captioning and visual question answering. These results demonstrate the effectiveness of combining targeted data and alignment objectives for enhancing fine-grained visual reasoning in MLLMs.

cs.CV↗

arXiv2025

ChemAU: Harness the Reasoning of LLMs in Chemical Research with Adaptive Uncertainty Estimation

Large Language Models (LLMs) are widely used across various scenarios due to their exceptional reasoning capabilities and natural language understanding. While LLMs demonstrate strong performance in tasks involving mathematics and coding, their effectiveness diminishes significantly when applied to chemistry-related problems. Chemistry problems typically involve long and complex reasoning steps, which contain specific terminology, including specialized symbol systems and complex nomenclature conventions. These characteristics often cause general LLMs to experience hallucinations during the reasoning process due to their lack of specific knowledge. However, existing methods are struggling to effectively leverage chemical expertise and formulas. Moreover, current uncertainty estimation methods, designed to mitigate potential reasoning errors, are unable to precisely identify specific steps or key knowledge. In this work, we propose a novel framework called ChemAU, which incorporates our adaptive uncertainty estimation method that applies different uncertainty values based on the position of reasoning steps within the whole reasoning chain. Leveraging this method, ChemAU identifies gaps in chemistry knowledge and precisely supplements chemical expertise with the specialized domain model, thereby correcting and updating the previously flawed reasoning chain. Our experiments with three popular LLMs across three chemistry datasets demonstrate that ChemAU significantly enhances both reasoning accuracy and uncertainty estimation.

cs.AI↗

arXiv2024

SemiCD-VL: Visual-Language Model Guidance Makes Better Semi-supervised Change Detector

Change Detection (CD) aims to identify pixels with semantic changes between images. However, annotating massive numbers of pixel-level images is labor-intensive and costly, especially for multi-temporal images, which require pixel-wise comparisons by human experts. Considering the excellent performance of visual language models (VLMs) for zero-shot, open-vocabulary, etc. with prompt-based reasoning, it is promising to utilize VLMs to make better CD under limited labeled data. In this paper, we propose a VLM guidance-based semi-supervised CD method, namely SemiCD-VL. The insight of SemiCD-VL is to synthesize free change labels using VLMs to provide additional supervision signals for unlabeled data. However, almost all current VLMs are designed for single-temporal images and cannot be directly applied to bi- or multi-temporal images. Motivated by this, we first propose a VLM-based mixed change event generation (CEG) strategy to yield pseudo labels for unlabeled CD data. Since the additional supervised signals provided by these VLM-driven pseudo labels may conflict with the pseudo labels from the consistency regularization paradigm (e.g. FixMatch), we propose the dual projection head for de-entangling different signal sources. Further, we explicitly decouple the bi-temporal images semantic representation through two auxiliary segmentation decoders, which are also guided by VLM. Finally, to make the model more adequately capture change representations, we introduce metric-aware supervision by feature-level contrastive loss in auxiliary branches. Extensive experiments show the advantage of SemiCD-VL. For instance, SemiCD-VL improves the FixMatch baseline by +5.3 IoU on WHU-CD and by +2.4 IoU on LEVIR-CD with 5% labels. In addition, our CEG strategy, in an un-supervised manner, can achieve performance far superior to state-of-the-art un-supervised CD methods.

cs.CV↗

arXiv2024

MV-CC: Mask Enhanced Video Model for Remote Sensing Change Caption

Remote sensing image change caption (RSICC) aims to provide natural language descriptions for bi-temporal remote sensing images. Since Change Caption (CC) task requires both spatial and temporal features, previous works follow an encoder-fusion-decoder architecture. They use an image encoder to extract spatial features and the fusion module to integrate spatial features and extract temporal features, which leads to increasingly complex manual design of the fusion module. In this paper, we introduce a novel video model-based paradigm without design of the fusion module and propose a Mask-enhanced Video model for Change Caption (MV-CC). Specifically, we use the off-the-shelf video encoder to simultaneously extract the temporal and spatial features of bi-temporal images. Furthermore, the types of changes in the CC are set based on specific task requirements, and to enable the model to better focus on the regions of interest, we employ masks obtained from the Change Detection (CD) method to explicitly guide the CC model. Experimental results demonstrate that our proposed method can obtain better performance compared with other state-of-the-art RSICC methods. The code is available at https://github.com/liuruixun/MV-CC.

cs.CV↗

arXiv2024

MBPU: A Plug-and-Play State Space Model for Point Cloud Upsamping with Fast Point Rendering

The task of point cloud upsampling (PCU) is to generate dense and uniform point clouds from sparse input captured by 3D sensors like LiDAR, holding potential applications in real yet is still a challenging task. Existing deep learning-based methods have shown significant achievements in this field. However, they still face limitations in effectively handling long sequences and addressing the issue of shrinkage artifacts around the surface of the point cloud. Inspired by the newly proposed Mamba, in this paper, we introduce a network named MBPU built on top of the Mamba architecture, which performs well in long sequence modeling, especially for large-scale point cloud upsampling, and achieves fast convergence speed. Moreover, MBPU is an arbitrary-scale upsampling framework as the predictor of point distance in the point refinement phase. At the same time, we simultaneously predict the 3D position shift and 1D point-to-point distance as regression quantities to constrain the global features while ensuring the accuracy of local details. We also introduce a fast differentiable renderer to further enhance the fidelity of the upsampled point cloud and reduce artifacts. It is noted that, by the merits of our fast point rendering, MBPU yields high-quality upsampled point clouds by effectively eliminating surface noise. Extensive experiments have demonstrated that our MBPU outperforms other off-the-shelf methods in terms of point cloud upsampling, especially for large-scale point clouds.

cs.CV↗

arXiv2024

On the Benefits of Traffic "Reprofiling'' -- The Single Hop Case

The need to guarantee hard delay bounds to traffic flows with deterministic traffic profiles, e.g., token buckets, arises in a number of network settings. Of interest are solutions that offer such guarantees while minimizing network bandwidth. The paper explores a basic building block towards realizing such solutions, namely, a single hop configuration. The main results are in the form of optimal solutions for meeting local deadlines under schedulers of varying complexity and therefore cost. The results demonstrate how judiciously modifying flows' traffic profiles, i.e., reprofiling them, can help simple schedulers reduce the bandwidth they require, often performing nearly as well as more complex ones.

cs.NI↗