Kimi-VL Technical Report

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), this model exhibits strong long-horizon reasoning capabilities. It achieves scores of 61.7 on MMMU, 36.8 on MathVision, and 71.3 on MathVista while maintaining the compact 2.8B activated LLM parameters, setting a new standard for efficient multimodal thinking models. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

Transferable Mask Transformer: Cross-domain Semantic Segmentation with Region-adaptive Transferability Estimation

Recent advances in Vision Transformers (ViTs) have set new benchmarks in semantic segmentation. However, when adapting pretrained ViTs to new target domains, significant performance degradation often occurs due to distribution shifts, resulting in suboptimal global attention. Since self-attention mechanisms are inherently data-driven, they may fail to effectively attend to key objects when source and target domains exhibit differences in texture, scale, or object co-occurrence patterns. While global and patch-level domain adaptation methods provide partial solutions, region-level adaptation with dynamically shaped regions is crucial due to spatial heterogeneity in transferability across different image areas. We present Transferable Mask Transformer (TMT), a novel region-level adaptation framework for semantic segmentation that aligns cross-domain representations through spatial transferability analysis. TMT consists of two key components: (1) An Adaptive Cluster-based Transferability Estimator (ACTE) that dynamically segments images into structurally and semantically coherent regions for localized transferability assessment, and (2) A Transferable Masked Attention (TMA) module that integrates region-specific transferability maps into ViTs' attention mechanisms, prioritizing adaptation in regions with low transferability and high semantic uncertainty. Comprehensive evaluations across 20 cross-domain pairs demonstrate TMT's superiority, achieving an average 2% MIoU improvement over vanilla fine-tuning and a 1.28% increase compared to state-of-the-art baselines. The source code will be publicly available.

SkyReels-A2: Compose Anything in Video Diffusion Transformers

This paper presents SkyReels-A2, a controllable video generation framework capable of assembling arbitrary visual elements (e.g., characters, objects, backgrounds) into synthesized videos based on textual prompts while maintaining strict consistency with reference images for each element. We term this task elements-to-video (E2V), whose primary challenges lie in preserving the fidelity of each reference element, ensuring coherent composition of the scene, and achieving natural outputs. To address these, we first design a comprehensive data pipeline to construct prompt-reference-video triplets for model training. Next, we propose a novel image-text joint embedding model to inject multi-element representations into the generative process, balancing element-specific consistency with global coherence and text alignment. We also optimize the inference pipeline for both speed and output stability. Moreover, we introduce a carefully curated benchmark for systematic evaluation, i.e, A2 Bench. Experiments demonstrate that our framework can generate diverse, high-quality videos with precise element control. SkyReels-A2 is the first open-source commercial grade model for the generation of E2V, performing favorably against advanced closed-source commercial models. We anticipate SkyReels-A2 will advance creative applications such as drama and virtual e-commerce, pushing the boundaries of controllable video generation.

ABC-GS: Alignment-Based Controllable Style Transfer for 3D Gaussian Splatting

3D scene stylization approaches based on Neural Radiance Fields (NeRF) achieve promising results by optimizing with Nearest Neighbor Feature Matching (NNFM) loss. However, NNFM loss does not consider global style information. In addition, the implicit representation of NeRF limits their fine-grained control over the resulting scenes. In this paper, we introduce ABC-GS, a novel framework based on 3D Gaussian Splatting to achieve high-quality 3D style transfer. To this end, a controllable matching stage is designed to achieve precise alignment between scene content and style features through segmentation masks. Moreover, a style transfer loss function based on feature alignment is proposed to ensure that the outcomes of style transfer accurately reflect the global style of the reference image. Furthermore, the original geometric information of the scene is preserved with the depth loss and Gaussian regularization terms. Extensive experiments show that our ABC-GS provides controllability of style transfer and achieves stylization results that are more faithfully aligned with the global style of the chosen artistic reference. Our homepage is available at https://vpx-ecnu.github.io/ABC-GS-website.

Analyzable Chain-of-Musical-Thought Prompting for High-Fidelity Music Generation

Autoregressive (AR) models have demonstrated impressive capabilities in generating high-fidelity music. However, the conventional next-token prediction paradigm in AR models does not align with the human creative process in music composition, potentially compromising the musicality of generated samples. To overcome this limitation, we introduce MusiCoT, a novel chain-of-thought (CoT) prompting technique tailored for music generation. MusiCoT empowers the AR model to first outline an overall music structure before generating audio tokens, thereby enhancing the coherence and creativity of the resulting compositions. By leveraging the contrastive language-audio pretraining (CLAP) model, we establish a chain of "musical thoughts", making MusiCoT scalable and independent of human-labeled data, in contrast to conventional CoT methods. Moreover, MusiCoT allows for in-depth analysis of music structure, such as instrumental arrangements, and supports music referencing -- accepting variable-length audio inputs as optional style references. This innovative approach effectively addresses copying issues, positioning MusiCoT as a vital practical method for music prompting. Our experimental results indicate that MusiCoT consistently achieves superior performance across both objective and subjective metrics, producing music quality that rivals state-of-the-art generation models. Our samples are available at https://MusiCoT.github.io/.

UltraFlwr -- An Efficient Federated Medical and Surgical Object Detection Framework

Object detection shows promise for medical and surgical applications such as cell counting and tool tracking. However, its faces multiple real-world edge deployment challenges including limited high-quality annotated data, data sharing restrictions, and computational constraints. In this work, we introduce UltraFlwr, a framework for federated medical and surgical object detection. By leveraging Federated Learning (FL), UltraFlwr enables decentralized model training across multiple sites without sharing raw data. To further enhance UltraFlwr's efficiency, we propose YOLO-PA, a set of novel Partial Aggregation (PA) strategies specifically designed for YOLO models in FL. YOLO-PA significantly reduces communication overhead by up to 83% per round while maintaining performance comparable to Full Aggregation (FA) strategies. Our extensive experiments on BCCD and m2cai16-tool-locations datasets demonstrate that YOLO-PA not only provides better client models compared to client-wise centralized training and FA strategies, but also facilitates efficient training and deployment across resource-constrained edge devices. Further, we also establish one of the first benchmarks in federated medical and surgical object detection. This paper advances the feasibility of training and deploying detection models on the edge, making federated object detection more practical for time-critical and resource-constrained medical and surgical applications. UltraFlwr is publicly available at https://github.com/KCL-BMEIS/UltraFlwr.

SpaceVLLM: Endowing Multimodal Large Language Model with Spatio-Temporal Video Grounding Capability

Multimodal large language models (MLLMs) have made remarkable progress in either temporal or spatial localization. However, they struggle to perform spatio-temporal video grounding. This limitation stems from two major challenges. Firstly, it is difficult to extract accurate spatio-temporal information of each frame in the video. Secondly, the substantial number of visual tokens makes it challenging to precisely map visual tokens of each frame to their corresponding spatial coordinates. To address these issues, we introduce SpaceVLLM, a MLLM endowed with spatio-temporal video grounding capability. Specifically, we adopt a set of interleaved Spatio-Temporal Aware Queries to capture temporal perception and dynamic spatial information. Moreover, we propose a Query-Guided Space Decoder to establish a corresponding connection between the queries and spatial coordinates. Additionally, due to the lack of spatio-temporal datasets, we construct the Unified Spatio-Temporal Grounding (Uni-STG) dataset, comprising 480K instances across three tasks. This dataset fully exploits the potential of MLLM to simultaneously facilitate localization in both temporal and spatial dimensions. Extensive experiments demonstrate that SpaceVLLM achieves the state-of-the-art performance across 11 benchmarks covering temporal, spatial, spatio-temporal and video understanding tasks, highlighting the effectiveness of our approach. Our code, datasets and model will be released.

Post-interactive Multimodal Trajectory Prediction for Autonomous Driving

Modeling the interactions among agents for trajectory prediction of autonomous driving has been challenging due to the inherent uncertainty in agents' behavior. The interactions involved in the predicted trajectories of agents, also called post-interactions, have rarely been considered in trajectory prediction models. To this end, we propose a coarse-to-fine Transformer for multimodal trajectory prediction, i.e., Pioformer, which explicitly extracts the post-interaction features to enhance the prediction accuracy. Specifically, we first build a Coarse Trajectory Network to generate coarse trajectories based on the observed trajectories and lane segments, in which the low-order interaction features are extracted with the graph neural networks. Next, we build a hypergraph neural network-based Trajectory Proposal Network to generate trajectory proposals, where the high-order interaction features are learned by the hypergraphs. Finally, the trajectory proposals are sent to the Proposal Refinement Network for further refinement. The observed trajectories and trajectory proposals are concatenated together as the inputs of the Proposal Refinement Network, in which the post-interaction features are learned by combining the previous interaction features and trajectory consistency features. Moreover, we propose a three-stage training scheme to facilitate the learning process. Extensive experiments on the Argoverse 1 dataset demonstrate the superiority of our method. Compared with the baseline HiVT-64, our model has reduced the prediction errors by 4.4%, 8.4%, 14.4%, 5.7% regarding metrics minADE6, minFDE6, MR6, and brier-minFDE6, respectively.

EVE: Towards End-to-End Video Subtitle Extraction with Vision-Language Models

The advent of Large Vision-Language Models (LVLMs) has advanced the video-based tasks, such as video captioning and video understanding. Some previous research indicates that taking texts in videos as input can further improve the performance of video understanding. As a type of indispensable information in short videos or movies, subtitles can assist LVLMs to better understand videos. Most existing methods for video subtitle extraction are based on a multi-stage framework, handling each frame independently. They can hardly exploit the temporal information of videos. Although some LVLMs exhibit the robust OCR capability, predicting accurate timestamps for subtitle texts is still challenging. In this paper, we propose an End-to-end Video Subtitle Extraction method, called EVE, which consists of three modules: a vision encoder, an adapter module, and a large language model. To effectively compress the visual tokens from the vision encoder, we propose a novel adapter InterleavedVT to interleave two modalities. It contains a visual compressor and a textual region compressor. The proposed InterleavedVT exploits both the merits of average pooling and Q-Former in token compression. Taking the temporal information of videos into account, we introduce a sliding-window mechanism in the textual region compressor. To benchmark the video subtitle extraction task, we propose a large dataset ViSa including 2.5M videos. Extensive experiments on ViSa demonstrate that the proposed EVE can outperform existing open-sourced tools and LVLMs.

Rethinking Video Tokenization: A Conditioned Diffusion-based Approach

Video tokenizers, which transform videos into compact latent representations, are key to video generation. Existing video tokenizers are based on the VAE architecture and follow a paradigm where an encoder compresses videos into compact latents, and a deterministic decoder reconstructs the original videos from these latents. In this paper, we propose a novel \underline{\textbf{C}}onditioned \underline{\textbf{D}}iffusion-based video \underline{\textbf{T}}okenizer entitled \textbf{\ourmethod}, which departs from previous methods by replacing the deterministic decoder with a 3D causal diffusion model. The reverse diffusion generative process of the decoder is conditioned on the latent representations derived via the encoder. With a feature caching and sampling acceleration, the framework efficiently reconstructs high-fidelity videos of arbitrary lengths. Results show that {\ourmethod} achieves state-of-the-art performance in video reconstruction tasks using just a single-step sampling. Even a smaller version of {\ourmethod} still achieves reconstruction results on par with the top two baselines. Furthermore, the latent video generation model trained using {\ourmethod} also shows superior performance.