Qiaole Dong

Qiaole Dong, Chenjie Cao, Yanwei Fu

Image inpainting has made significant advances in recent years. However, it is still challenging to recover corrupted images with both vivid textures and reasonable structures. Some specific methods only tackle regular textures while losing holistic structures due to the limited receptive fields of convolutional neural networks (CNNs). On the other hand, attention-based models can learn better long-range dependency for the structure recovery, but they are limited by the heavy computation for inference with large image sizes. To address these issues, we propose to leverage an additional structure restorer to facilitate the image inpainting incrementally. The proposed model restores holistic image structures with a powerful attention-based transformer model in a fixed low-resolution sketch space. Such a grayscale space is easy to be upsampled to larger scales to convey correct structural information. Our structure restorer can be integrated with other pretrained inpainting models efficiently with the zero-initialized residual addition. Furthermore, a masking positional encoding strategy is utilized to improve the performance with large irregular masks. Extensive experiments on various datasets validate the efficacy of our model compared with other competitors. Our codes are released in https://github.com/DQiaole/ZITS_inpainting.

Qiaole Dong, Chenjie Cao, Yanwei Fu

Optical flow estimation is a challenging problem remaining unsolved. Recent deep learning based optical flow models have achieved considerable success. However, these models often train networks from the scratch on standard optical flow data, which restricts their ability to robustly and geometrically match image features. In this paper, we propose a rethinking to previous optical flow estimation. We particularly leverage Geometric Image Matching (GIM) as a pre-training task for the optical flow estimation (MatchFlow) with better feature representations, as GIM shares some common challenges as optical flow estimation, and with massive labeled real-world data. Thus, matching static scenes helps to learn more fundamental feature correlations of objects and scenes with consistent displacements. Specifically, the proposed MatchFlow model employs a QuadTree attention-based network pre-trained on MegaDepth to extract coarse features for further flow regression. Extensive experiments show that our model has great cross-dataset generalization. Our method achieves 11.5% and 10.1% error reduction from GMA on Sintel clean pass and KITTI test set. At the time of anonymous submission, our MatchFlow(G) enjoys state-of-the-art performance on Sintel clean and final pass compared to published approaches with comparable computation and memory footprint. Codes and models will be released in https://github.com/DQiaole/MatchFlow.

Chenjie Cao, Qiaole Dong, Yanwei Fu

Image inpainting involves filling missing areas of a corrupted image. Despite impressive results have been achieved recently, restoring images with both vivid textures and reasonable structures remains a significant challenge. Previous methods have primarily addressed regular textures while disregarding holistic structures due to the limited receptive fields of Convolutional Neural Networks (CNNs). To this end, we study learning a Zero-initialized residual addition based Incremental Transformer on Structural priors (ZITS++), an improved model upon our conference work, ZITS. Specifically, given one corrupt image, we present the Transformer Structure Restorer (TSR) module to restore holistic structural priors at low image resolution, which are further upsampled by Simple Structure Upsampler (SSU) module to higher image resolution. To recover image texture details, we use the Fourier CNN Texture Restoration (FTR) module, which is strengthened by Fourier and large-kernel attention convolutions. Furthermore, to enhance the FTR, the upsampled structural priors from TSR are further processed by Structure Feature Encoder (SFE) and optimized with the Zero-initialized Residual Addition (ZeroRA) incrementally. Besides, a new masking positional encoding is proposed to encode the large irregular masks. Compared with ZITS, ZITS++ improves the FTR's stability and inpainting ability with several techniques. More importantly, we comprehensively explore the effects of various image priors for inpainting and investigate how to utilize them to address high-resolution image inpainting with extensive experiments. This investigation is orthogonal to most inpainting approaches and can thus significantly benefit the community. Codes and models will be released in https://github.com/ewrfcas/ZITS-PlusPlus.

Chenjie Cao, Qiaole Dong, Yanwei Fu

Many recent inpainting works have achieved impressive results by leveraging Deep Neural Networks (DNNs) to model various prior information for image restoration. Unfortunately, the performance of these methods is largely limited by the representation ability of vanilla Convolutional Neural Networks (CNNs) backbones.On the other hand, Vision Transformers (ViT) with self-supervised pre-training have shown great potential for many visual recognition and object detection tasks. A natural question is whether the inpainting task can be greatly benefited from the ViT backbone? However, it is nontrivial to directly replace the new backbones in inpainting networks, as the inpainting is an inverse problem fundamentally different from the recognition tasks. To this end, this paper incorporates the pre-training based Masked AutoEncoder (MAE) into the inpainting model, which enjoys richer informative priors to enhance the inpainting process. Moreover, we propose to use attention priors from MAE to make the inpainting model learn more long-distance dependencies between masked and unmasked regions. Sufficient ablations have been discussed about the inpainting and the self-supervised pre-training models in this paper. Besides, experiments on both Places2 and FFHQ demonstrate the effectiveness of our proposed model. Codes and pre-trained models are released in https://github.com/ewrfcas/MAE-FAR.

Qiaole Dong, Bo Zhao, Yanwei Fu

Recently, Google proposes DDVM which for the first time demonstrates that a general diffusion model for image-to-image translation task works impressively well on optical flow estimation task without any specific designs like RAFT. However, DDVM is still a closed-source model with the expensive and private Palette-style pretraining. In this technical report, we present the first open-source DDVM by reproducing it. We study several design choices and find those important ones. By training on 40k public data with 4 GPUs, our reproduction achieves comparable performance to the closed-source DDVM. The code and model have been released in https://github.com/DQiaole/FlowDiffusion_pytorch.

Yun Wang, Junjie Hu, Qiaole Dong et al.

Temporally consistent depth estimation from stereo video is critical for real-world applications such as augmented reality, where inconsistent depth estimation disrupts the immersion of users. Despite its importance, this task remains challenging due to the difficulty in modeling long-term temporal consistency in a computationally efficient manner. Previous methods attempt to address this by aggregating spatio-temporal information but face a fundamental trade-off: limited temporal modeling provides only modest gains, whereas capturing long-range dependencies significantly increases computational cost. To address this limitation, we introduce a memory buffer for modeling long-range spatio-temporal consistency while achieving efficient dynamic stereo matching. Inspired by the two-stage decision-making process in humans, we propose a \textbf{P}ick-and-\textbf{P}lay \textbf{M}emory (PPM) construction module for dynamic \textbf{Stereo} matching, dubbed as \textbf{PPMStereo}. PPM consists of a `pick' process that identifies the most relevant frames and a `play' process that weights the selected frames adaptively for spatio-temporal aggregation. This two-stage collaborative process maintains a compact yet highly informative memory buffer while achieving temporally consistent information aggregation. Extensive experiments validate the effectiveness of PPMStereo, demonstrating state-of-the-art performance in both accuracy and temporal consistency. % Notably, PPMStereo achieves 0.62/1.11 TEPE on the Sintel clean/final (17.3\% \& 9.02\% improvements over BiDAStereo) with fewer computational costs. Codes are available at \textcolor{blue}{https://github.com/cocowy1/PPMStereo}.

Zhihao Zhang, Qiaole Dong, Qi Zhang et al.

Post-training algorithms such as Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are widely used to adapt multimodal large language models to downstream tasks. While effective at task adaptation, their impact on prior knowledge remains unclear. In this paper, we introduce jigsaw puzzles as a novel task absent from existing pretraining corpora and systematically study the behavior of SFT and RFT on open-source multimodal model, Qwen2.5-VL series. Our experiments reveal a sharp trade-off: SFT enables rapid task acquisition but leads to catastrophic forgetting, whereas RFT learns more slowly but maintains prior knowledge. We study this phenomenon through learning dynamics by examining both the magnitude and direction of how training data influence prior knowledge. Our analysis shows that RFT mainly reinforces correct samples naturally aligned with the base model's probability landscape, leading to weaker interference with prior knowledge. Moreover, training on RFT-simulated rollouts, which exert a small magnitude of influence and are well aligned in direction to prior knowledge, allows SFT to preserve prior knowledge better while rapidly learning new tasks. These findings suggest that distribution of training data, rather than algorithmic differences, plays a central role in forgetting, and highlight RFT's potential for stable continual learning in multimodal large language models.

Chenjie Cao, Yunuo Cai, Qiaole Dong et al.

This paper introduces LeftRefill, an innovative approach to efficiently harness large Text-to-Image (T2I) diffusion models for reference-guided image synthesis. As the name implies, LeftRefill horizontally stitches reference and target views together as a whole input. The reference image occupies the left side, while the target canvas is positioned on the right. Then, LeftRefill paints the right-side target canvas based on the left-side reference and specific task instructions. Such a task formulation shares some similarities with contextual inpainting, akin to the actions of a human painter. This novel formulation efficiently learns both structural and textured correspondence between reference and target without other image encoders or adapters. We inject task and view information through cross-attention modules in T2I models, and further exhibit multi-view reference ability via the re-arranged self-attention modules. These enable LeftRefill to perform consistent generation as a generalized model without requiring test-time fine-tuning or model modifications. Thus, LeftRefill can be seen as a simple yet unified framework to address reference-guided synthesis. As an exemplar, we leverage LeftRefill to address two different challenges: reference-guided inpainting and novel view synthesis, based on the pre-trained StableDiffusion. Codes and models are released at https://github.com/ewrfcas/LeftRefill.

Qiaole Dong, Yanwei Fu

Optical flow is a classical task that is important to the vision community. Classical optical flow estimation uses two frames as input, whilst some recent methods consider multiple frames to explicitly model long-range information. The former ones limit their ability to fully leverage temporal coherence along the video sequence; and the latter ones incur heavy computational overhead, typically not possible for real-time flow estimation. Some multi-frame-based approaches even necessitate unseen future frames for current estimation, compromising real-time applicability in safety-critical scenarios. To this end, we present MemFlow, a real-time method for optical flow estimation and prediction with memory. Our method enables memory read-out and update modules for aggregating historical motion information in real-time. Furthermore, we integrate resolution-adaptive re-scaling to accommodate diverse video resolutions. Besides, our approach seamlessly extends to the future prediction of optical flow based on past observations. Leveraging effective historical motion aggregation, our method outperforms VideoFlow with fewer parameters and faster inference speed on Sintel and KITTI-15 datasets in terms of generalization performance. At the time of submission, MemFlow also leads in performance on the 1080p Spring dataset. Codes and models will be available at: https://dqiaole.github.io/MemFlow/.

Mingqi Wu, Zhihao Zhang, Qiaole Dong et al.

Reasoning in large language models has long been a central research focus, and recent studies employing reinforcement learning (RL) have introduced diverse methods that yield substantial performance gains with minimal or even no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance performance. However, these breakthroughs are predominantly observed for the mathematically strong Qwen2.5 series on benchmarks such as MATH-500, AMC, and AIME, and seldom transfer to models like Llama, which warrants a more in-depth investigation. In this work, our empirical analysis reveals that pre-training on massive web-scale corpora leaves Qwen2.5 susceptible to data contamination in widely used benchmarks. Consequently, conclusions derived from contaminated benchmarks on Qwen2.5 series may be unreliable. To obtain trustworthy evaluation results, we introduce a generator that creates fully clean arithmetic problems of arbitrary length and difficulty, dubbed RandomCalculation. Using this leakage-free dataset, we show that only accurate reward signals yield steady improvements that surpass the base model's performance boundary in mathematical reasoning, whereas random or incorrect rewards do not. Moreover, we conduct more fine-grained analyses to elucidate the factors underlying the different performance observed on the MATH-500 and RandomCalculation benchmarks. Consequently, we recommend that future studies evaluate models on uncontaminated benchmarks and, when feasible, test various model series to ensure trustworthy conclusions about RL and related methods.

Yikai Wang, Chenjie Cao, Ke Fan et al.

Current image manipulation primarily centers on static manipulation, such as replacing specific regions within an image or altering its overall style. In this paper, we introduce an innovative dynamic manipulation task, subject repositioning. This task involves relocating a user-specified subject to a desired position while preserving the image's fidelity. Our research reveals that the fundamental sub-tasks of subject repositioning, which include filling the void left by the repositioned subject, reconstructing obscured portions of the subject and blending the subject to be consistent with surrounding areas, can be effectively reformulated as a unified, prompt-guided inpainting task. Consequently, we can employ a single diffusion generative model to address these sub-tasks using various task prompts learned through our proposed task inversion technique. Additionally, we integrate pre-processing and post-processing techniques to further enhance the quality of subject repositioning. These elements together form our SEgment-gEnerate-and-bLEnd (SEELE) framework. To assess SEELE's effectiveness in subject repositioning, we assemble a real-world subject repositioning dataset called ReS. Results of SEELE on ReS demonstrate its efficacy. Code and ReS dataset are available at https://yikai-wang.github.io/seele/.

Bingwen Zhu, Yuqian Fu, Qiaole Dong et al.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in vision-language understanding. Yet, human perception is inherently multisensory, integrating sight, sound, and motion to reason about the world. Among these modalities, sound provides indispensable cues about spatial layout, off-screen events, and causal interactions, particularly in egocentric settings where auditory and visual signals are tightly coupled. To this end, we introduce EgoSound, the first benchmark designed to systematically evaluate egocentric sound understanding in MLLMs. EgoSound unifies data from Ego4D and EgoBlind, encompassing both sighted and sound-dependent experiences. It defines a seven-task taxonomy spanning intrinsic sound perception, spatial localization, causal inference, and cross-modal reasoning. Constructed through a multi-stage auto-generative pipeline, EgoSound contains 7315 validated QA pairs across 900 videos. Comprehensive experiments on nine state-of-the-art MLLMs reveal that current models exhibit emerging auditory reasoning abilities but remain limited in fine-grained spatial and causal understanding. EgoSound establishes a challenging foundation for advancing multisensory egocentric intelligence, bridging the gap between seeing and truly hearing the world.

Ming Xie, Junqiu Yu, Qiaole Dong et al.

Recent image-to-video (I2V) based video inpainting methods have made significant strides by leveraging single-image priors and modeling temporal consistency across masked frames. Nevertheless, these methods suffer from severe content degradation within video chunks. Furthermore, the absence of a robust frame alignment scheme compromises intra-chunk and inter-chunk spatiotemporal stability, resulting in insufficient control over the entire video. To address these limitations, we propose VidPivot, a novel framework that decouples video inpainting into two sub-tasks: multi-frame consistent image inpainting and masked area motion propagation. Our approach introduces frame interval priors as spatiotemporal cues to guide the inpainting process. To enhance cross-frame coherence, we design a FrameProp Module that implements a frame content propagation strategy, diffusing reference frame content into subsequent frames via a splicing mechanism. Additionally, a dedicated context controller encodes these coherent frame priors into the I2V generative backbone, effectively serving as soft constrain to suppress content distortion during generation. Extensive evaluations demonstrate that VidPivot achieves competitive performance across diverse benchmarks and generalizes well to different video inpainting scenarios.

Qiaole Dong, Yanwei Fu

Dense point tracking is a challenging task requiring the continuous tracking of every point in the initial frame throughout a substantial portion of a video, even in the presence of occlusions. Traditional methods use optical flow models to directly estimate long-range motion, but they often suffer from appearance drifting without considering temporal consistency. Recent point tracking algorithms usually depend on sliding windows for indirect information propagation from the first frame to the current one, which is slow and less effective for long-range tracking. To account for temporal consistency and enable efficient information propagation, we present a lightweight and fast model with \textbf{S}treaming memory for dense \textbf{PO}int \textbf{T}racking and online video processing. The \textbf{SPOT} framework features three core components: a customized memory reading module for feature enhancement, a sensory memory for short-term motion dynamics modeling, and a visibility-guided splatting module for accurate information propagation. This combination enables SPOT to perform dense point tracking with state-of-the-art accuracy on the CVO benchmark, as well as comparable or superior performance to offline models on sparse tracking benchmarks such as TAP-Vid and RoboTAP. Notably, SPOT with 10$\times$ smaller parameter numbers operates at least 2$\times$ faster than previous state-of-the-art models while maintaining the best performance on CVO. We will release the models and codes at: https://dqiaole.github.io/SPOT/.