KVShare: Semantic-Aware Key-Value Cache Sharing for Efficient Large Language Model Inference

This paper presents KVShare, a multi-user Key-Value (KV) Cache sharing technology based on semantic similarity, designed to enhance the inference efficiency of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs). Addressing the limitations of existing prefix caching (strict text prefix matching) and semantic caching (loss of response diversity), KVShare achieves fine-grained KV cache reuse through semantic alignment algorithms and differential editing operations. Experiments on real-world user conversation datasets demonstrate that KVShare improves KV cache hit rates by over 60%, while maintaining output quality comparable to full computation (no significant degradation in BLEU and Rouge-L metrics). This approach effectively reduces GPU resource consumption and is applicable to scenarios with repetitive queries, such as healthcare and education.

Vamba: Understanding Hour-Long Videos with Hybrid Mamba-Transformers

State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames (640$\times$360) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least 50% reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by 4.3% on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.

Accelerating Video Diffusion Models via Distribution Matching

Generative models, particularly diffusion models, have made significant success in data synthesis across various modalities, including images, videos, and 3D assets. However, current diffusion models are computationally intensive, often requiring numerous sampling steps that limit their practical application, especially in video generation. This work introduces a novel framework for diffusion distillation and distribution matching that dramatically reduces the number of inference steps while maintaining-and potentially improving-generation quality. Our approach focuses on distilling pre-trained diffusion models into a more efficient few-step generator, specifically targeting video generation. By leveraging a combination of video GAN loss and a novel 2D score distribution matching loss, we demonstrate the potential to generate high-quality video frames with substantially fewer sampling steps. To be specific, the proposed method incorporates a denoising GAN discriminator to distil from the real data and a pre-trained image diffusion model to enhance the frame quality and the prompt-following capabilities. Experimental results using AnimateDiff as the teacher model showcase the method's effectiveness, achieving superior performance in just four sampling steps compared to existing techniques.

Long Video Diffusion Generation with Segmented Cross-Attention and Content-Rich Video Data Curation

We introduce Presto, a novel video diffusion model designed to generate 15-second videos with long-range coherence and rich content. Extending video generation methods to maintain scenario diversity over long durations presents significant challenges. To address this, we propose a Segmented Cross-Attention (SCA) strategy, which splits hidden states into segments along the temporal dimension, allowing each segment to cross-attend to a corresponding sub-caption. SCA requires no additional parameters, enabling seamless incorporation into current DiT-based architectures. To facilitate high-quality long video generation, we build the LongTake-HD dataset, consisting of 261k content-rich videos with scenario coherence, annotated with an overall video caption and five progressive sub-captions. Experiments show that our Presto achieves 78.5% on the VBench Semantic Score and 100% on the Dynamic Degree, outperforming existing state-of-the-art video generation methods. This demonstrates that our proposed Presto significantly enhances content richness, maintains long-range coherence, and captures intricate textual details. More details are displayed on our project page: https://presto-video.github.io/.

VISTA: Enhancing Long-Duration and High-Resolution Video Understanding by Video Spatiotemporal Augmentation

Current large multimodal models (LMMs) face significant challenges in processing and comprehending long-duration or high-resolution videos, which is mainly due to the lack of high-quality datasets. To address this issue from a data-centric perspective, we propose VISTA, a simple yet effective Video Spatiotemporal Augmentation framework that synthesizes long-duration and high-resolution video instruction-following pairs from existing video-caption datasets. VISTA spatially and temporally combines videos to create new synthetic videos with extended durations and enhanced resolutions, and subsequently produces question-answer pairs pertaining to these newly synthesized videos. Based on this paradigm, we develop seven video augmentation methods and curate VISTA-400K, a video instruction-following dataset aimed at enhancing long-duration and high-resolution video understanding. Finetuning various video LMMs on our data resulted in an average improvement of 3.3% across four challenging benchmarks for long-video understanding. Furthermore, we introduce the first comprehensive high-resolution video understanding benchmark HRVideoBench, on which our finetuned models achieve a 6.5% performance gain. These results highlight the effectiveness of our framework.

Fleximo: Towards Flexible Text-to-Human Motion Video Generation

Current methods for generating human motion videos rely on extracting pose sequences from reference videos, which restricts flexibility and control. Additionally, due to the limitations of pose detection techniques, the extracted pose sequences can sometimes be inaccurate, leading to low-quality video outputs. We introduce a novel task aimed at generating human motion videos solely from reference images and natural language. This approach offers greater flexibility and ease of use, as text is more accessible than the desired guidance videos. However, training an end-to-end model for this task requires millions of high-quality text and human motion video pairs, which are challenging to obtain. To address this, we propose a new framework called Fleximo, which leverages large-scale pre-trained text-to-3D motion models. This approach is not straightforward, as the text-generated skeletons may not consistently match the scale of the reference image and may lack detailed information. To overcome these challenges, we introduce an anchor point based rescale method and design a skeleton adapter to fill in missing details and bridge the gap between text-to-motion and motion-to-video generation. We also propose a video refinement process to further enhance video quality. A large language model (LLM) is employed to decompose natural language into discrete motion sequences, enabling the generation of motion videos of any desired length. To assess the performance of Fleximo, we introduce a new benchmark called MotionBench, which includes 400 videos across 20 identities and 20 motions. We also propose a new metric, MotionScore, to evaluate the accuracy of motion following. Both qualitative and quantitative results demonstrate that our method outperforms existing text-conditioned image-to-video generation methods. All code and model weights will be made publicly available.

Improving Multi-Subject Consistency in Open-Domain Image Generation with Isolation and Reposition Attention

Training-free diffusion models have achieved remarkable progress in generating multi-subject consistent images within open-domain scenarios. The key idea of these methods is to incorporate reference subject information within the attention layer. However, existing methods still obtain suboptimal performance when handling numerous subjects. This paper reveals the two primary issues contributing to this deficiency. Firstly, there is undesired interference among different subjects within the target image. Secondly, tokens tend to reference nearby tokens, which reduces the effectiveness of the attention mechanism when there is a significant positional difference between subjects in reference and target images. To address these challenges, we propose a training-free diffusion model with Isolation and Reposition Attention, named IR-Diffusion. Specifically, Isolation Attention ensures that multiple subjects in the target image do not reference each other, effectively eliminating the subject fusion. On the other hand, Reposition Attention involves scaling and repositioning subjects in both reference and target images to the same position within the images. This ensures that subjects in the target image can better reference those in the reference image, thereby maintaining better consistency. Extensive experiments demonstrate that the proposed methods significantly enhance multi-subject consistency, outperforming all existing methods in open-domain scenarios.

Allegro: Open the Black Box of Commercial-Level Video Generation Model

Significant advancements have been made in the field of video generation, with the open-source community contributing a wealth of research papers and tools for training high-quality models. However, despite these efforts, the available information and resources remain insufficient for achieving commercial-level performance. In this report, we open the black box and introduce $\textbf{Allegro}$, an advanced video generation model that excels in both quality and temporal consistency. We also highlight the current limitations in the field and present a comprehensive methodology for training high-performance, commercial-level video generation models, addressing key aspects such as data, model architecture, training pipeline, and evaluation. Our user study shows that Allegro surpasses existing open-source models and most commercial models, ranking just behind Hailuo and Kling. Code: https://github.com/rhymes-ai/Allegro , Model: https://huggingface.co/rhymes-ai/Allegro , Gallery: https://rhymes.ai/allegro_gallery .

Self-Supervised Learning of Deviation in Latent Representation for Co-speech Gesture Video Generation

Gestures are pivotal in enhancing co-speech communication. While recent works have mostly focused on point-level motion transformation or fully supervised motion representations through data-driven approaches, we explore the representation of gestures in co-speech, with a focus on self-supervised representation and pixel-level motion deviation, utilizing a diffusion model which incorporates latent motion features. Our approach leverages self-supervised deviation in latent representation to facilitate hand gestures generation, which are crucial for generating realistic gesture videos. Results of our first experiment demonstrate that our method enhances the quality of generated videos, with an improvement from 2.7 to 4.5% for FGD, DIV, and FVD, and 8.1% for PSNR, 2.5% for SSIM over the current state-of-the-art methods.

A First Look At Efficient And Secure On-Device LLM Inference Against KV Leakage

Running LLMs on end devices has garnered significant attention recently due to their advantages in privacy preservation. With the advent of lightweight LLM models and specially designed GPUs, on-device LLM inference has achieved the necessary accuracy and performance metrics. However, we have identified that LLM inference on GPUs can leak privacy-sensitive intermediate information, specifically the KV pairs. An attacker could exploit these KV pairs to reconstruct the entire user conversation, leading to significant vulnerabilities. Existing solutions, such as Fully Homomorphic Encryption (FHE) and Trusted Execution Environments (TEE), are either too computation-intensive or resource-limited. To address these issues, we designed KV-Shield, which operates in two phases. In the initialization phase, it permutes the weight matrices so that all KV pairs are correspondingly permuted. During the runtime phase, the attention vector is inversely permuted to ensure the correctness of the layer output. All permutation-related operations are executed within the TEE, ensuring that insecure GPUs cannot access the original KV pairs, thus preventing conversation reconstruction. Finally, we theoretically analyze the correctness of KV-Shield, along with its advantages and overhead.