Daquan Zhou

Lin Xu, Zhiyuan Hu, Daquan Zhou et al. · berkeley

Large Language Models (LLMs) have significantly advanced natural language processing, demonstrating exceptional reasoning, tool usage, and memory capabilities. As their applications expand into multi-agent environments, there arises a need for a comprehensive evaluation framework that captures LLMs' reasoning, planning, collaboration, and other social abilities. This work introduces a novel competition-based benchmark framework specifically designed to assess LLMs within multi-agent settings, providing quantitative metrics to evaluate their judgment, reasoning, deception, self-awareness, cooperation, coordination, and rationality. We utilize two social deduction games alongside three game-theory scenarios to create diverse environments. Our frame is fortified with the probabilistic graphic modeling (PGM) method, enhancing the LLMs' capabilities in navigating complex social and cognitive dimensions. We evaluate seven LLMs, quantitatively highlighting a significant capability gap of over threefold between the strongest, GPT o1, and the weakest, Llama-2-70B. It also confirms that our PGM enhancement boosts the abilities of all selected models by an average of 37%. Our data and code can be found here https://github.com/cathyxl/MAgIC.

Dongze Lian, Daquan Zhou, Jiashi Feng et al.

Existing fine-tuning methods either tune all parameters of the pre-trained model (full fine-tuning), which is not efficient, or only tune the last linear layer (linear probing), which suffers a significant accuracy drop compared to the full fine-tuning. In this paper, we propose a new parameter-efficient fine-tuning method termed as SSF, representing that researchers only need to Scale and Shift the deep Features extracted by a pre-trained model to catch up with the performance of full fine-tuning. In this way, SSF also surprisingly outperforms other parameter-efficient fine-tuning approaches even with a smaller number of tunable parameters. Furthermore, different from some existing parameter-efficient fine-tuning methods (e.g., Adapter or VPT) that introduce the extra parameters and computational cost in the training and inference stages, SSF only adds learnable parameters during the training stage, and these additional parameters can be merged into the original pre-trained model weights via re-parameterization in the inference phase. With the proposed SSF, our model obtains 2.46% (90.72% vs. 88.54%) and 11.48% (73.10% vs. 65.57%) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy compared to the full fine-tuning but only fine-tuning about 0.3M parameters. We also conduct amounts of experiments in various model families (CNNs, Transformers, and MLPs) and datasets. Results on 26 image classification datasets in total and 3 robustness & out-of-distribution datasets show the effectiveness of SSF. Code is available at https://github.com/dongzelian/SSF.

Daquan Zhou, Zhiding Yu, Enze Xie et al.

Recent studies show that Vision Transformers(ViTs) exhibit strong robustness against various corruptions. Although this property is partly attributed to the self-attention mechanism, there is still a lack of systematic understanding. In this paper, we examine the role of self-attention in learning robust representations. Our study is motivated by the intriguing properties of the emerging visual grouping in Vision Transformers, which indicates that self-attention may promote robustness through improved mid-level representations. We further propose a family of fully attentional networks (FANs) that strengthen this capability by incorporating an attentional channel processing design. We validate the design comprehensively on various hierarchical backbones. Our model achieves a state-of-the-art 87.1% accuracy and 35.8% mCE on ImageNet-1k and ImageNet-C with 76.8M parameters. We also demonstrate state-of-the-art accuracy and robustness in two downstream tasks: semantic segmentation and object detection. Code is available at: https://github.com/NVlabs/FAN.

Xingyi Yang, Daquan Zhou, Songhua Liu et al.

In this paper, we explore a novel knowledge-transfer task, termed as Deep Model Reassembly (DeRy), for general-purpose model reuse. Given a collection of heterogeneous models pre-trained from distinct sources and with diverse architectures, the goal of DeRy, as its name implies, is to first dissect each model into distinctive building blocks, and then selectively reassemble the derived blocks to produce customized networks under both the hardware resource and performance constraints. Such ambitious nature of DeRy inevitably imposes significant challenges, including, in the first place, the feasibility of its solution. We strive to showcase that, through a dedicated paradigm proposed in this paper, DeRy can be made not only possibly but practically efficiently. Specifically, we conduct the partitions of all pre-trained networks jointly via a cover set optimization, and derive a number of equivalence set, within each of which the network blocks are treated as functionally equivalent and hence interchangeable. The equivalence sets learned in this way, in turn, enable picking and assembling blocks to customize networks subject to certain constraints, which is achieved via solving an integer program backed up with a training-free proxy to estimate the task performance. The reassembled models, give rise to gratifying performances with the user-specified constraints satisfied. We demonstrate that on ImageNet, the best reassemble model achieves 78.6% top-1 accuracy without fine-tuning, which could be further elevated to 83.2% with end-to-end training. Our code is available at https://github.com/Adamdad/DeRy

Yifan Zhang, Daquan Zhou, Bryan Hooi et al.

The power of DNNs relies heavily on the quantity and quality of training data. However, collecting and annotating data on a large scale is often expensive and time-consuming. To address this issue, we explore a new task, termed dataset expansion, aimed at expanding a ready-to-use small dataset by automatically creating new labeled samples. To this end, we present a Guided Imagination Framework (GIF) that leverages cutting-edge generative models like DALL-E2 and Stable Diffusion (SD) to "imagine" and create informative new data from the input seed data. Specifically, GIF conducts data imagination by optimizing the latent features of the seed data in the semantically meaningful space of the prior model, resulting in the creation of photo-realistic images with new content. To guide the imagination towards creating informative samples for model training, we introduce two key criteria, i.e., class-maintained information boosting and sample diversity promotion. These criteria are verified to be essential for effective dataset expansion: GIF-SD obtains 13.5% higher model accuracy on natural image datasets than unguided expansion with SD. With these essential criteria, GIF successfully expands small datasets in various scenarios, boosting model accuracy by 36.9% on average over six natural image datasets and by 13.5% on average over three medical datasets. The source code is available at https://github.com/Vanint/DatasetExpansion.

Ziheng Qin, Kai Wang, Zangwei Zheng et al.

Data pruning aims to obtain lossless performances with less overall cost. A common approach is to filter out samples that make less contribution to the training. This could lead to gradient expectation bias compared to the original data. To solve this problem, we propose \textbf{InfoBatch}, a novel framework aiming to achieve lossless training acceleration by unbiased dynamic data pruning. Specifically, InfoBatch randomly prunes a portion of less informative samples based on the loss distribution and rescales the gradients of the remaining samples to approximate the original gradient. As a plug-and-play and architecture-agnostic framework, InfoBatch consistently obtains lossless training results on classification, semantic segmentation, vision pertaining, and instruction fine-tuning tasks. On CIFAR10/100, ImageNet-1K, and ADE20K, InfoBatch losslessly saves 40\% overall cost. For pertaining MAE and diffusion model, InfoBatch can respectively save 24.8\% and 27\% cost. For LLaMA instruction fine-tuning, InfoBatch is also able to save 20\% cost and is compatible with coreset selection methods. The code is publicly available at \href{https://github.com/henryqin1997/InfoBatch}{github.com/NUS-HPC-AI-Lab/InfoBatch}.

Enze Xie, Zhiding Yu, Daquan Zhou et al.

In this paper, we propose M$^2$BEV, a unified framework that jointly performs 3D object detection and map segmentation in the Birds Eye View~(BEV) space with multi-camera image inputs. Unlike the majority of previous works which separately process detection and segmentation, M$^2$BEV infers both tasks with a unified model and improves efficiency. M$^2$BEV efficiently transforms multi-view 2D image features into the 3D BEV feature in ego-car coordinates. Such BEV representation is important as it enables different tasks to share a single encoder. Our framework further contains four important designs that benefit both accuracy and efficiency: (1) An efficient BEV encoder design that reduces the spatial dimension of a voxel feature map. (2) A dynamic box assignment strategy that uses learning-to-match to assign ground-truth 3D boxes with anchors. (3) A BEV centerness re-weighting that reinforces with larger weights for more distant predictions, and (4) Large-scale 2D detection pre-training and auxiliary supervision. We show that these designs significantly benefit the ill-posed camera-based 3D perception tasks where depth information is missing. M$^2$BEV is memory efficient, allowing significantly higher resolution images as input, with faster inference speed. Experiments on nuScenes show that M$^2$BEV achieves state-of-the-art results in both 3D object detection and BEV segmentation, with the best single model achieving 42.5 mAP and 57.0 mIoU in these two tasks, respectively.

Yu-Huan Wu, Shi-Chen Zhang, Yun Liu et al.

Semantic segmentation tasks naturally require high-resolution information for pixel-wise segmentation and global context information for class prediction. While existing vision transformers demonstrate promising performance, they often utilize high-resolution context modeling, resulting in a computational bottleneck. In this work, we challenge conventional wisdom and introduce the Low-Resolution Self-Attention (LRSA) mechanism to capture global context at a significantly reduced computational cost, i.e., FLOPs. Our approach involves computing self-attention in a fixed low-resolution space regardless of the input image's resolution, with additional 3x3 depth-wise convolutions to capture fine details in the high-resolution space. We demonstrate the effectiveness of our LRSA approach by building the LRFormer, a vision transformer with an encoder-decoder structure. Extensive experiments on the ADE20K, COCO-Stuff, and Cityscapes datasets demonstrate that LRFormer outperforms state-of-the-art models. Code is available at https://github.com/yuhuan-wu/LRFormer.

Daquan Zhou, Weimin Wang, Hanshu Yan et al.

We present an efficient text-to-video generation framework based on latent diffusion models, termed MagicVideo. MagicVideo can generate smooth video clips that are concordant with the given text descriptions. Due to a novel and efficient 3D U-Net design and modeling video distributions in a low-dimensional space, MagicVideo can synthesize video clips with 256x256 spatial resolution on a single GPU card, which takes around 64x fewer computations than the Video Diffusion Models (VDM) in terms of FLOPs. In specific, unlike existing works that directly train video models in the RGB space, we use a pre-trained VAE to map video clips into a low-dimensional latent space and learn the distribution of videos' latent codes via a diffusion model. Besides, we introduce two new designs to adapt the U-Net denoiser trained on image tasks to video data: a frame-wise lightweight adaptor for the image-to-video distribution adjustment and a directed temporal attention module to capture temporal dependencies across frames. Thus, we can exploit the informative weights of convolution operators from a text-to-image model for accelerating video training. To ameliorate the pixel dithering in the generated videos, we also propose a novel VideoVAE auto-encoder for better RGB reconstruction. We conduct extensive experiments and demonstrate that MagicVideo can generate high-quality video clips with either realistic or imaginary content. Refer to \url{https://magicvideo.github.io/#} for more examples.

Xingyi Yang, Daquan Zhou, Jiashi Feng et al.

Despite the recent visually-pleasing results achieved, the massive computational cost has been a long-standing flaw for diffusion probabilistic models (DPMs), which, in turn, greatly limits their applications on resource-limited platforms. Prior methods towards efficient DPM, however, have largely focused on accelerating the testing yet overlooked their huge complexity and sizes. In this paper, we make a dedicated attempt to lighten DPM while striving to preserve its favourable performance. We start by training a small-sized latent diffusion model (LDM) from scratch, but observe a significant fidelity drop in the synthetic images. Through a thorough assessment, we find that DPM is intrinsically biased against high-frequency generation, and learns to recover different frequency components at different time-steps. These properties make compact networks unable to represent frequency dynamics with accurate high-frequency estimation. Towards this end, we introduce a customized design for slim DPM, which we term as Spectral Diffusion (SD), for light-weight image synthesis. SD incorporates wavelet gating in its architecture to enable frequency dynamic feature extraction at every reverse steps, and conducts spectrum-aware distillation to promote high-frequency recovery by inverse weighting the objective based on spectrum magni tudes. Experimental results demonstrate that, SD achieves 8-18x computational complexity reduction as compared to the latent diffusion models on a series of conditional and unconditional image generation tasks while retaining competitive image fidelity.

Kai Wang, Jianyang Gu, Daquan Zhou et al.

Dataset distillation reduces the network training cost by synthesizing small and informative datasets from large-scale ones. Despite the success of the recent dataset distillation algorithms, three drawbacks still limit their wider application: i). the synthetic images perform poorly on large architectures; ii). they need to be re-optimized when the distillation ratio changes; iii). the limited diversity restricts the performance when the distillation ratio is large. In this paper, we propose a novel distillation scheme to \textbf{D}istill information of large train sets \textbf{i}nto generative \textbf{M}odels, named DiM. Specifically, DiM learns to use a generative model to store the information of the target dataset. During the distillation phase, we minimize the differences in logits predicted by a models pool between real and generated images. At the deployment stage, the generative model synthesizes various training samples from random noises on the fly. Due to the simple yet effective designs, the trained DiM can be directly applied to different distillation ratios and large architectures without extra cost. We validate the proposed DiM across 4 datasets and achieve state-of-the-art results on all of them. To the best of our knowledge, we are the first to achieve higher accuracy on complex architectures than simple ones, such as 75.1\% with ResNet-18 and 72.6\% with ConvNet-3 on ten images per class of CIFAR-10. Besides, DiM outperforms previous methods with 10\% $\sim$ 22\% when images per class are 1 and 10 on the SVHN dataset.

Jiawei Du, Daquan Zhou, Jiashi Feng et al.

Modern deep neural networks (DNNs) have achieved state-of-the-art performances but are typically over-parameterized. The over-parameterization may result in undesirably large generalization error in the absence of other customized training strategies. Recently, a line of research under the name of Sharpness-Aware Minimization (SAM) has shown that minimizing a sharpness measure, which reflects the geometry of the loss landscape, can significantly reduce the generalization error. However, SAM-like methods incur a two-fold computational overhead of the given base optimizer (e.g. SGD) for approximating the sharpness measure. In this paper, we propose Sharpness-Aware Training for Free, or SAF, which mitigates the sharp landscape at almost zero additional computational cost over the base optimizer. Intuitively, SAF achieves this by avoiding sudden drops in the loss in the sharp local minima throughout the trajectory of the updates of the weights. Specifically, we suggest a novel trajectory loss, based on the KL-divergence between the outputs of DNNs with the current weights and past weights, as a replacement of the SAM's sharpness measure. This loss captures the rate of change of the training loss along the model's update trajectory. By minimizing it, SAF ensures the convergence to a flat minimum with improved generalization capabilities. Extensive empirical results show that SAF minimizes the sharpness in the same way that SAM does, yielding better results on the ImageNet dataset with essentially the same computational cost as the base optimizer.

Enze Xie, Lewei Yao, Han Shi et al.

Diffusion models have proven to be highly effective in generating high-quality images. However, adapting large pre-trained diffusion models to new domains remains an open challenge, which is critical for real-world applications. This paper proposes DiffFit, a parameter-efficient strategy to fine-tune large pre-trained diffusion models that enable fast adaptation to new domains. DiffFit is embarrassingly simple that only fine-tunes the bias term and newly-added scaling factors in specific layers, yet resulting in significant training speed-up and reduced model storage costs. Compared with full fine-tuning, DiffFit achieves 2$\times$ training speed-up and only needs to store approximately 0.12\% of the total model parameters. Intuitive theoretical analysis has been provided to justify the efficacy of scaling factors on fast adaptation. On 8 downstream datasets, DiffFit achieves superior or competitive performances compared to the full fine-tuning while being more efficient. Remarkably, we show that DiffFit can adapt a pre-trained low-resolution generative model to a high-resolution one by adding minimal cost. Among diffusion-based methods, DiffFit sets a new state-of-the-art FID of 3.02 on ImageNet 512$\times$512 benchmark by fine-tuning only 25 epochs from a public pre-trained ImageNet 256$\times$256 checkpoint while being 30$\times$ more training efficient than the closest competitor.

Yang Zhao, Zhijie Lin, Daquan Zhou et al.

LLMs have demonstrated remarkable abilities at interacting with humans through language, especially with the usage of instruction-following data. Recent advancements in LLMs, such as MiniGPT-4, LLaVA, and X-LLM, further enlarge their abilities by incorporating multi-modal inputs, including image, video, and speech. Despite their effectiveness at generating precise and detailed language understanding of the given modality signal, these LLMs give up the ability to ground specific parts of inputs, thus only constructing a coarse-grained mapping. However, explicit and informative correspondence between text and other modalities will not only improve the user experience but also help to expand the application scenario of multi-modal LLMs. Therefore, we propose BuboGPT, a multi-modal LLM with visual grounding that can perform cross-modal interaction between vision, audio and language, providing fine-grained understanding of visual objects and other given modalities. As a result, BuboGPT is able to point out the specific location of an object in the image, when it is generating response or description for that object. Our contributions are two-fold: 1) An off-the-shelf visual grounding module based on SAM that extracts entities in a sentence and find corresponding masks in the image. 2) A two-stage training scheme and instruction dataset to endow joint text-image-audio understanding. Our experiments show that BuboGPT achieves impressive multi-modality understanding and visual grounding abilities during the interaction with human. It performs consistently well when provided by arbitrary modality combinations (either aligned or unaligned). Our code, model and dataset are available at https://bubo-gpt.github.io .

Yupeng Zhou, Daquan Zhou, Zuo-Liang Zhu et al.

Recent advancements in diffusion models have showcased their impressive capacity to generate visually striking images. Nevertheless, ensuring a close match between the generated image and the given prompt remains a persistent challenge. In this work, we identify that a crucial factor leading to the text-image mismatch issue is the inadequate cross-modality relation learning between the prompt and the output image. To better align the prompt and image content, we advance the cross-attention with an adaptive mask, which is conditioned on the attention maps and the prompt embeddings, to dynamically adjust the contribution of each text token to the image features. This mechanism explicitly diminishes the ambiguity in semantic information embedding from the text encoder, leading to a boost of text-to-image consistency in the synthesized images. Our method, termed MaskDiffusion, is training-free and hot-pluggable for popular pre-trained diffusion models. When applied to the latent diffusion models, our MaskDiffusion can significantly improve the text-to-image consistency with negligible computation overhead compared to the original diffusion models.

Chenyu Wang, Zhen Dong, Daquan Zhou et al. · berkeley

The utilization of large-scale neural networks on Processing-In-Memory (PIM) accelerators encounters challenges due to constrained on-chip memory capacity. To tackle this issue, current works explore model compression algorithms to reduce the size of Convolutional Neural Networks (CNNs). Most of these algorithms either aim to represent neural operators with reduced-size parameters (e.g., quantization) or search for the best combinations of neural operators (e.g., neural architecture search). Designing neural operators to align with PIM accelerators' specifications is an area that warrants further study. In this paper, we introduce the Epitome, a lightweight neural operator offering convolution-like functionality, to craft memory-efficient CNN operators for PIM accelerators (EPIM). On the software side, we evaluate epitomes' latency and energy on PIM accelerators and introduce a PIM-aware layer-wise design method to enhance their hardware efficiency. We apply epitome-aware quantization to further reduce the size of epitomes. On the hardware side, we modify the datapath of current PIM accelerators to accommodate epitomes and implement a feature map reuse technique to reduce computation cost. Experimental results reveal that our 3-bit quantized EPIM-ResNet50 attains 71.59% top-1 accuracy on ImageNet, reducing crossbar areas by 30.65 times. EPIM surpasses the state-of-the-art pruning methods on PIM.

Jun Hao Liew, Hanshu Yan, Daquan Zhou et al.

Have you ever imagined what a corgi-alike coffee machine or a tiger-alike rabbit would look like? In this work, we attempt to answer these questions by exploring a new task called semantic mixing, aiming at blending two different semantics to create a new concept (e.g., corgi + coffee machine -- > corgi-alike coffee machine). Unlike style transfer, where an image is stylized according to the reference style without changing the image content, semantic blending mixes two different concepts in a semantic manner to synthesize a novel concept while preserving the spatial layout and geometry. To this end, we present MagicMix, a simple yet effective solution based on pre-trained text-conditioned diffusion models. Motivated by the progressive generation property of diffusion models where layout/shape emerges at early denoising steps while semantically meaningful details appear at later steps during the denoising process, our method first obtains a coarse layout (either by corrupting an image or denoising from a pure Gaussian noise given a text prompt), followed by injection of conditional prompt for semantic mixing. Our method does not require any spatial mask or re-training, yet is able to synthesize novel objects with high fidelity. To improve the mixing quality, we further devise two simple strategies to provide better control and flexibility over the synthesized content. With our method, we present our results over diverse downstream applications, including semantic style transfer, novel object synthesis, breed mixing, and concept removal, demonstrating the flexibility of our method. More results can be found on the project page https://magicmix.github.io

Daquan Zhou, Kai Wang, Jianyang Gu et al.

State-of-the-art deep neural networks are trained with large amounts (millions or even billions) of data. The expensive computation and memory costs make it difficult to train them on limited hardware resources, especially for recent popular large language models (LLM) and computer vision models (CV). Recent popular dataset distillation methods are thus developed, aiming to reduce the number of training samples via synthesizing small-scale datasets via gradient matching. However, as the gradient calculation is coupled with the specific network architecture, the synthesized dataset is biased and performs poorly when used for training unseen architectures. To address these limitations, we present dataset quantization (DQ), a new framework to compress large-scale datasets into small subsets which can be used for training any neural network architectures. Extensive experiments demonstrate that DQ is able to generate condensed small datasets for training unseen network architectures with state-of-the-art compression ratios for lossless model training. To the best of our knowledge, DQ is the first method that can successfully distill large-scale datasets such as ImageNet-1k with a state-of-the-art compression ratio. Notably, with 60% data from ImageNet and 20% data from Alpaca's instruction tuning data, the models can be trained with negligible or no performance drop for both vision tasks (including classification, semantic segmentation, and object detection) as well as language tasks (including instruction tuning tasks such as BBH and DROP).

Yuxuan Chen, Changwei Lv, Yunduo Xiao et al.

Large Language Models (LLMs) are central to the one-for-all intelligent paradigm, but they face a fundamental challenge when dealing with heterogeneous scientific data such as molecules: the inherent gap between discrete linguistic symbols and topological molecular or continuous reaction data leads to significant information loss and semantic noise in text-based reasoning. We propose SciCore-Mol, a modular framework that bridges this gap through three deeply integrated pluggable cognitive modules: a topology-aware perception module, a latent diffusion-based molecular generation module, and a reaction-aware reasoning module. Each module is coupled to the LLM backbone through learned representation interfaces, enabling richer information exchange than is possible with text-only tool feedback. Our experiments on diverse chemical tasks demonstrate that SciCore-Mol achieves strong comprehensive performance across molecular understanding, generation, reaction prediction, and general chemistry knowledge, with an 8B-parameter open-source system that is competitive with and in several dimensions surpasses proprietary large models. This work provides a systematic blueprint for equipping LLMs with scientific expertise through decoupled, pluggable, and flexibly orchestrated modules, with direct implications for drug design, chemical synthesis, and broader scientific discovery.

Yilin Zhao, Xinbin Yuan, Shanghua Gao et al.

In this technical report, we target generating anthropomorphized personas for LLM-based characters in an online manner, including visual appearance, personality and tones, with only text descriptions. To achieve this, we first leverage the in-context learning capability of LLMs for personality generation by carefully designing a set of system prompts. We then propose two novel concepts: the mixture of voices (MoV) and the mixture of diffusers (MoD) for diverse voice and appearance generation. For MoV, we utilize the text-to-speech (TTS) algorithms with a variety of pre-defined tones and select the most matching one based on the user-provided text description automatically. For MoD, we combine the recent popular text-to-image generation techniques and talking head algorithms to streamline the process of generating talking objects. We termed the whole framework as ChatAnything. With it, users could be able to animate anything with any personas that are anthropomorphic using just a few text inputs. However, we have observed that the anthropomorphic objects produced by current generative models are often undetectable by pre-trained face landmark detectors, leading to failure of the face motion generation, even if these faces possess human-like appearances because those images are nearly seen during the training (e.g., OOD samples). To address this issue, we incorporate pixel-level guidance to infuse human face landmarks during the image generation phase. To benchmark these metrics, we have built an evaluation dataset. Based on it, we verify that the detection rate of the face landmark is significantly increased from 57.0% to 92.5% thus allowing automatic face animation based on generated speech content. The code and more results can be found at https://chatanything.github.io/.

Lin Xu, Yilin Zhao, Daquan Zhou et al.

Vision-language pre-training has significantly elevated performance across a wide range of image-language applications. Yet, the pre-training process for video-related tasks demands exceptionally large computational and data resources, which hinders the progress of video-language models. This paper investigates a straight-forward, highly efficient, and resource-light approach to adapting an existing image-language pre-trained model for dense video understanding. Our preliminary experiments reveal that directly fine-tuning pre-trained image-language models with multiple frames as inputs on video datasets leads to performance saturation or even a drop. Our further investigation reveals that it is largely attributed to the bias of learned high-norm visual features. Motivated by this finding, we propose a simple but effective pooling strategy to smooth the feature distribution along the temporal dimension and thus reduce the dominant impacts from the extreme features. The new model is termed Pooling LLaVA, or PLLaVA in short. PLLaVA achieves new state-of-the-art performance on modern benchmark datasets for both video question-answer and captioning tasks. Notably, on the recent popular VideoChatGPT benchmark, PLLaVA achieves a score of 3.48 out of 5 on average of five evaluated dimensions, exceeding the previous SOTA results from GPT4V (IG-VLM) by 9%. On the latest multi-choice benchmark MVBench, PLLaVA achieves 58.1% accuracy on average across 20 sub-tasks, 14.5% higher than GPT4V (IG-VLM). Code is available at https://pllava.github.io/

Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng et al.

For recent diffusion-based generative models, maintaining consistent content across a series of generated images, especially those containing subjects and complex details, presents a significant challenge. In this paper, we propose a new way of self-attention calculation, termed Consistent Self-Attention, that significantly boosts the consistency between the generated images and augments prevalent pretrained diffusion-based text-to-image models in a zero-shot manner. To extend our method to long-range video generation, we further introduce a novel semantic space temporal motion prediction module, named Semantic Motion Predictor. It is trained to estimate the motion conditions between two provided images in the semantic spaces. This module converts the generated sequence of images into videos with smooth transitions and consistent subjects that are significantly more stable than the modules based on latent spaces only, especially in the context of long video generation. By merging these two novel components, our framework, referred to as StoryDiffusion, can describe a text-based story with consistent images or videos encompassing a rich variety of contents. The proposed StoryDiffusion encompasses pioneering explorations in visual story generation with the presentation of images and videos, which we hope could inspire more research from the aspect of architectural modifications. Our code is made publicly available at https://github.com/HVision-NKU/StoryDiffusion.

Enze Xie, Junsong Chen, Yuyang Zhao et al.

This paper presents SANA-1.5, a linear Diffusion Transformer for efficient scaling in text-to-image generation. Building upon SANA-1.0, we introduce three key innovations: (1) Efficient Training Scaling: A depth-growth paradigm that enables scaling from 1.6B to 4.8B parameters with significantly reduced computational resources, combined with a memory-efficient 8-bit optimizer. (2) Model Depth Pruning: A block importance analysis technique for efficient model compression to arbitrary sizes with minimal quality loss. (3) Inference-time Scaling: A repeated sampling strategy that trades computation for model capacity, enabling smaller models to match larger model quality at inference time. Through these strategies, SANA-1.5 achieves a text-image alignment score of 0.81 on GenEval, which can be further improved to 0.96 through inference scaling with VILA-Judge, establishing a new SoTA on GenEval benchmark. These innovations enable efficient model scaling across different compute budgets while maintaining high quality, making high-quality image generation more accessible. Our code and pre-trained models are released.

Yufan Deng, Zilin Pan, Hongyu Zhang et al.

Video generation models have significantly advanced embodied intelligence, unlocking new possibilities for generating diverse robot data that capture perception, reasoning, and action in the physical world. However, synthesizing high-quality videos that accurately reflect real-world robotic interactions remains challenging, and the lack of a standardized benchmark limits fair comparisons and progress. To address this gap, we introduce a comprehensive robotics benchmark, RBench, designed to evaluate robot-oriented video generation across five task domains and four distinct embodiments. It assesses both task-level correctness and visual fidelity through reproducible sub-metrics, including structural consistency, physical plausibility, and action completeness. Evaluation of 25 representative models highlights significant deficiencies in generating physically realistic robot behaviors. Furthermore, the benchmark achieves a Spearman correlation coefficient of 0.96 with human evaluations, validating its effectiveness. While RBench provides the necessary lens to identify these deficiencies, achieving physical realism requires moving beyond evaluation to address the critical shortage of high-quality training data. Driven by these insights, we introduce a refined four-stage data pipeline, resulting in RoVid-X, the largest open-source robotic dataset for video generation with 4 million annotated video clips, covering thousands of tasks and enriched with comprehensive physical property annotations. Collectively, this synergistic ecosystem of evaluation and data establishes a robust foundation for rigorous assessment and scalable training of video models, accelerating the evolution of embodied AI toward general intelligence.

Kewei Zhang, Jin Wang, Sensen Gao et al.

End-to-end autonomous driving via Vision-Language-Action (VLA) models demands a precarious balance between high-fidelity trajectory planning and efficient inference. Existing paradigms typically fall short: autoregressive (AR) VLAs are memory-bandwidth-bound on edge hardware and prone to exposure-bias drift, while full-sequence diffusion models preclude KV-cache reuse and suffer from "logical leakage" that violates the fundamental perceive-then-plan causality. We present Fast-dDrive, a block-diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across them. Leveraging the observation that driving VLAs often emit structured JSON-like outputs, Fast-dDrive freezes structural tokens into a section scaffold and employs a section-aware training recipe that prioritizes safety-critical planning. We further introduce Scaffold Speculative Decoding to achieve AR-equivalent quality at significantly higher throughput. Finally, we propose a low-overhead test-time scaling scheme: by forking $N$ stochastic trajectory rollouts from a single shared-prefix KV cache and averaging them, we effectively suppress prediction variance at a fractional computational cost. Empirical results demonstrate that Fast-dDrive redefines the speed-accuracy frontier for driving agents. On the WOD-E2E test set, Fast-dDrive achieves SOTA ADE@3s and ADE@5s, alongside the highest RFS among diffusion-based VLAs; on nuScenes, it reduces average L2 error to $0.32$m (a $22\%$ improvement). When integrated with SGLang, our framework delivers $12\times$ throughput speedup over the AR baseline, narrowing the gap between high-capacity VLAs and the efficiency demands of real-time on-vehicle deployment.

Ze Ma, Daquan Zhou, Chun-Hsiao Yeh et al. · berkeley

Creating content with specified identities (ID) has attracted significant interest in the field of generative models. In the field of text-to-image generation (T2I), subject-driven creation has achieved great progress with the identity controlled via reference images. However, its extension to video generation is not well explored. In this work, we propose a simple yet effective subject identity controllable video generation framework, termed Video Custom Diffusion (VCD). With a specified identity defined by a few images, VCD reinforces the identity characteristics and injects frame-wise correlation at the initialization stage for stable video outputs. To achieve this, we propose three novel components that are essential for high-quality identity preservation and stable video generation: 1) a noise initialization method with 3D Gaussian Noise Prior for better inter-frame stability; 2) an ID module based on extended Textual Inversion trained with the cropped identity to disentangle the ID information from the background 3) Face VCD and Tiled VCD modules to reinforce faces and upscale the video to higher resolution while preserving the identity's features. We conducted extensive experiments to verify that VCD is able to generate stable videos with better ID over the baselines. Besides, with the transferability of the encoded identity in the ID module, VCD is also working well with personalized text-to-image models available publicly. The codes are available at https://github.com/Zhen-Dong/Magic-Me.

Libo Zhang, Zekun Li, Tianyu Li et al.

Humans exhibit adaptive, context-sensitive responses to egocentric visual input. However, faithfully modeling such reactions from egocentric video remains challenging due to the dual requirements of strictly causal generation and precise 3D spatial alignment. To tackle this problem, we first construct the Human Reaction Dataset (HRD) to address data scarcity and misalignment by building a spatially aligned egocentric video-reaction dataset, as existing datasets (e.g., ViMo) suffer from significant spatial inconsistency between the egocentric video and reaction motion, e.g., dynamically moving motions are always paired with fixed-camera videos. Leveraging HRD, we present EgoReAct, the first autoregressive framework that generates 3D-aligned human reaction motions from egocentric video streams in real-time. We first compress the reaction motion into a compact yet expressive latent space via a Vector Quantised-Variational AutoEncoder and then train a Generative Pre-trained Transformer for reaction generation from the visual input. EgoReAct incorporates 3D dynamic features, i.e., metric depth, and head dynamics during the generation, which effectively enhance spatial grounding. Extensive experiments demonstrate that EgoReAct achieves remarkably higher realism, spatial consistency, and generation efficiency compared with prior methods, while maintaining strict causality during generation. We will release code, models, and data upon acceptance.

Hongyu Zhang, Yufan Deng, Zilin Pan et al.

Generating high-quality videos from complex temporal descriptions that contain multiple sequential actions is a key unsolved problem. Existing methods are constrained by an inherent trade-off: using multiple short prompts fed sequentially into the model improves action fidelity but compromises temporal consistency, while a single complex prompt preserves consistency at the cost of prompt-following capability. We attribute this problem to two primary causes: 1) temporal misalignment between video content and the prompt, and 2) conflicting attention coupling between motion-related visual objects and their associated text conditions. To address these challenges, we propose a novel, training-free attention mechanism, Temporal-wise Separable Attention (TS-Attn), which dynamically rearranges attention distribution to ensure temporal awareness and global coherence in multi-event scenarios. TS-Attn can be seamlessly integrated into various pre-trained text-to-video models, boosting StoryEval-Bench scores by 33.5% and 16.4% on Wan2.1-T2V-14B and Wan2.2-T2V-A14B with only a 2% increase in inference time. It also supports plug-and-play usage across models for multi-event image-to-video generation. The source code and project page are available at https://github.com/Hong-yu-Zhang/TS-Attn.

Haolin Qin, Daquan Zhou, Tingfa Xu et al.

Transformers have astounding representational power but typically consume considerable computation which is quadratic with image resolution. The prevailing Swin transformer reduces computational costs through a local window strategy. However, this strategy inevitably causes two drawbacks: (1) the local window-based self-attention hinders global dependency modeling capability; (2) recent studies point out that local windows impair robustness. To overcome these challenges, we pursue a preferable trade-off between computational cost and performance. Accordingly, we propose a novel factorization self-attention mechanism (FaSA) that enjoys both the advantages of local window cost and long-range dependency modeling capability. By factorizing the conventional attention matrix into sparse sub-attention matrices, FaSA captures long-range dependencies while aggregating mixed-grained information at a computational cost equivalent to the local window-based self-attention. Leveraging FaSA, we present the factorization vision transformer (FaViT) with a hierarchical structure. FaViT achieves high performance and robustness, with linear computational complexity concerning input image spatial resolution. Extensive experiments have shown FaViT's advanced performance in classification and downstream tasks. Furthermore, it also exhibits strong model robustness to corrupted and biased data and hence demonstrates benefits in favor of practical applications. In comparison to the baseline model Swin-T, our FaViT-B2 significantly improves classification accuracy by 1% and robustness by 7%, while reducing model parameters by 14%. Our code will soon be publicly available at https://github.com/q2479036243/FaViT.

Kewei Zhang, Ye Huang, Yufan Deng et al.

While the Transformer architecture dominates many fields, its quadratic self-attention complexity hinders its use in large-scale applications. Linear attention offers an efficient alternative, but its direct application often degrades performance, with existing fixes typically re-introducing computational overhead through extra modules (e.g., depthwise separable convolution) that defeat the original purpose. In this work, we identify a key failure mode in these methods: global context collapse, where the model loses representational diversity. To address this, we propose Multi-Head Linear Attention (MHLA), which preserves this diversity by computing attention within divided heads along the token dimension. We prove that MHLA maintains linear complexity while recovering much of the expressive power of softmax attention, and verify its effectiveness across multiple domains, achieving a 3.6\% improvement on ImageNet classification, a 6.3\% gain on NLP, a 12.6\% improvement on image generation, and a 41\% enhancement on video generation under the same time complexity.

Yiyang Fu, Chubin Zhang, Shukai Gong et al.

It is infeasible to encompass all possible disturbances within the training dataset. This raises a critical question regarding the robustness of Vision-Language-Action (VLA) models when encountering unseen real-world visual disturbances, particularly under imperfect visual conditions. In this work, we conduct a systematic study based on recent state-of-the-art VLA models and reveal a significant performance drop when visual disturbances absent from the training data are introduced. To mitigate this issue, we propose a lightweight adapter module grounded in information theory, termed the Information Bottleneck Adapter (IB-Adapter), which selectively filters potential noise from visual inputs. Without requiring any extra data or augmentation strategies, IB-Adapter consistently improves over the baseline by an average of 30%, while adding fewer than 10M parameters, demonstrating notable efficiency and effectiveness. Furthermore, even with a 14x smaller backbone (0.5B parameters) and no pre-training on the Open X-Embodiment dataset, our model StableVLA achieves robustness competitive with 7B-scale state-of-the-art VLAs. With negligible parameter overhead (<10M), our approach maintains accuracy on long-horizon tasks and surpasses OpenPi under both synthetic and physical visual corruptions.

Hongwei Yi, Shitong Shao, Tian Ye et al.

In this technical report, we present Magic 1-For-1 (Magic141), an efficient video generation model with optimized memory consumption and inference latency. The key idea is simple: factorize the text-to-video generation task into two separate easier tasks for diffusion step distillation, namely text-to-image generation and image-to-video generation. We verify that with the same optimization algorithm, the image-to-video task is indeed easier to converge over the text-to-video task. We also explore a bag of optimization tricks to reduce the computational cost of training the image-to-video (I2V) models from three aspects: 1) model convergence speedup by using a multi-modal prior condition injection; 2) inference latency speed up by applying an adversarial step distillation, and 3) inference memory cost optimization with parameter sparsification. With those techniques, we are able to generate 5-second video clips within 3 seconds. By applying a test time sliding window, we are able to generate a minute-long video within one minute with significantly improved visual quality and motion dynamics, spending less than 1 second for generating 1 second video clips on average. We conduct a series of preliminary explorations to find out the optimal tradeoff between computational cost and video quality during diffusion step distillation and hope this could be a good foundation model for open-source explorations. The code and the model weights are available at https://github.com/DA-Group-PKU/Magic-1-For-1.

Shitong Shao, Hongwei Yi, Hanzhong Guo et al.

Recently, open-source video diffusion models (VDMs), such as WanX, Magic141 and HunyuanVideo, have been scaled to over 10 billion parameters. These large-scale VDMs have demonstrated significant improvements over smaller-scale VDMs across multiple dimensions, including enhanced visual quality and more natural motion dynamics. However, these models face two major limitations: (1) High inference overhead: Large-scale VDMs require approximately 10 minutes to synthesize a 28-step video on a single H100 GPU. (2) Limited in portrait video synthesis: Models like WanX-I2V and HunyuanVideo-I2V often produce unnatural facial expressions and movements in portrait videos. To address these challenges, we propose MagicDistillation, a novel framework designed to reduce inference overhead while ensuring the generalization of VDMs for portrait video synthesis. Specifically, we primarily use sufficiently high-quality talking video to fine-tune Magic141, which is dedicated to portrait video synthesis. We then employ LoRA to effectively and efficiently fine-tune the fake DiT within the step distillation framework known as distribution matching distillation (DMD). Following this, we apply weak-to-strong (W2S) distribution matching and minimize the discrepancy between the fake data distribution and the ground truth distribution, thereby improving the visual fidelity and motion dynamics of the synthesized videos. Experimental results on portrait video synthesis demonstrate the effectiveness of MagicDistillation, as our method surpasses Euler, LCM, and DMD baselines in both FID/FVD metrics and VBench. Moreover, MagicDistillation, requiring only 4 steps, also outperforms WanX-I2V (14B) and HunyuanVideo-I2V (13B) on visualization and VBench. Our project page is https://magicdistillation.github.io/MagicDistillation/.

Hanzhong Guo, Hongwei Yi, Daquan Zhou et al.

Latent diffusion models have made great strides in generating expressive portrait videos with accurate lip-sync and natural motion from a single reference image and audio input. However, these models are far from real-time, often requiring many sampling steps that take minutes to generate even one second of video-significantly limiting practical use. We introduce OSA-LCM (One-Step Avatar Latent Consistency Model), paving the way for real-time diffusion-based avatars. Our method achieves comparable video quality to existing methods but requires only one sampling step, making it more than 10x faster. To accomplish this, we propose a novel avatar discriminator design that guides lip-audio consistency and motion expressiveness to enhance video quality in limited sampling steps. Additionally, we employ a second-stage training architecture using an editing fine-tuned method (EFT), transforming video generation into an editing task during training to effectively address the temporal gap challenge in single-step generation. Experiments demonstrate that OSA-LCM outperforms existing open-source portrait video generation models while operating more efficiently with a single sampling step.

Quan-Sheng Zeng, Yunheng Li, Daquan Zhou et al.

Open-vocabulary image segmentation has been advanced through the synergy between mask generators and vision-language models like Contrastive Language-Image Pre-training (CLIP). Previous approaches focus on generating masks while aligning mask features with text embeddings during training. In this paper, we observe that relying on generated low-quality masks can weaken the alignment of vision and language in regional representations. This motivates us to present a new fine-tuning framework, named MaskCLIP++, which uses ground-truth masks instead of generated masks to enhance the mask classification capability of CLIP. Due to the limited diversity of image segmentation datasets with mask annotations, we propose incorporating a consistency alignment principle during fine-tuning, which alleviates categorical bias toward the fine-tuning dataset. After low-cost fine-tuning, MaskCLIP++ significantly improves the mask classification performance on multi-domain datasets. Combining with the mask generator in previous state-of-the-art mask-based open vocabulary segmentation methods, we achieve performance improvements of +1.7, +2.3, +2.1, +3.1, and +0.3 mIoU on the A-847, PC-459, A-150, PC-59, and PAS-20 datasets, respectively. Code is avaliable at https://github.com/HVision-NKU/MaskCLIPpp .

Yufan Deng, Daquan Zhou

Progress in embodied intelligence increasingly depends on scalable data infrastructure. While vision and language have scaled with internet corpora, learning physical interaction remains constrained by the lack of large, diverse, and richly annotated human activity data. We present HumanNet, a one-million-hour human-centric video corpus that captures how humans interact with the physical world at scale. HumanNet spans both first-person and third-person perspectives and covers fine-grained activities, human-object interactions, tool use, and long-horizon behaviors across diverse real-world environments. Beyond raw video, the dataset provides interaction-centric annotations, including captions, motion descriptions, and hand and body-related signals, enabling motion-aware and interaction-aware learning. Beyond scale, HumanNet introduces a systematic data curation paradigm for embodied learning, where human-centric filtering, temporal structuring, viewpoint diversity, and annotation enrichment are treated as first-class design principles. This design transforms unstructured internet video into a scalable substrate for representation learning, activity understanding, motion generation, and human-to-robot transfer. We conduct a first-step validation on the value of this design through controlled vision-language-action ablation: under a fixed set of validation data, continued training from the Qwen VLM model with 1000 hours of egocentric video drawn from HumanNet surpasses the continued training with 100 hours of real-robot data from Magic Cobot, indicating that egocentric human video could be a scalable and cost-effective substitute for robot data. By building this project, we aim to explore the opportunity to scale embodied foundation models using human-centric videos, rather than relying solely on robot-specific data.

Sucheng Ren, Daquan Zhou, Shengfeng He et al.

Recent Vision Transformer~(ViT) models have demonstrated encouraging results across various computer vision tasks, thanks to their competence in modeling long-range dependencies of image patches or tokens via self-attention. These models, however, usually designate the similar receptive fields of each token feature within each layer. Such a constraint inevitably limits the ability of each self-attention layer in capturing multi-scale features, thereby leading to performance degradation in handling images with multiple objects of different scales. To address this issue, we propose a novel and generic strategy, termed shunted self-attention~(SSA), that allows ViTs to model the attentions at hybrid scales per attention layer. The key idea of SSA is to inject heterogeneous receptive field sizes into tokens: before computing the self-attention matrix, it selectively merges tokens to represent larger object features while keeping certain tokens to preserve fine-grained features. This novel merging scheme enables the self-attention to learn relationships between objects with different sizes and simultaneously reduces the token numbers and the computational cost. Extensive experiments across various tasks demonstrate the superiority of SSA. Specifically, the SSA-based transformer achieves 84.0\% Top-1 accuracy and outperforms the state-of-the-art Focal Transformer on ImageNet with only half of the model size and computation cost, and surpasses Focal Transformer by 1.3 mAP on COCO and 2.9 mIOU on ADE20K under similar parameter and computation cost. Code has been released at https://github.com/OliverRensu/Shunted-Transformer.

Zihang Jiang, Qibin Hou, Li Yuan et al.

In this paper, we present token labeling -- a new training objective for training high-performance vision transformers (ViTs). Different from the standard training objective of ViTs that computes the classification loss on an additional trainable class token, our proposed one takes advantage of all the image patch tokens to compute the training loss in a dense manner. Specifically, token labeling reformulates the image classification problem into multiple token-level recognition problems and assigns each patch token with an individual location-specific supervision generated by a machine annotator. Experiments show that token labeling can clearly and consistently improve the performance of various ViT models across a wide spectrum. For a vision transformer with 26M learnable parameters serving as an example, with token labeling, the model can achieve 84.4% Top-1 accuracy on ImageNet. The result can be further increased to 86.4% by slightly scaling the model size up to 150M, delivering the minimal-sized model among previous models (250M+) reaching 86%. We also show that token labeling can clearly improve the generalization capability of the pre-trained models on downstream tasks with dense prediction, such as semantic segmentation. Our code and all the training details will be made publicly available at https://github.com/zihangJiang/TokenLabeling.

Daquan Zhou, Bingyi Kang, Xiaojie Jin et al.

Vision transformers (ViTs) have been successfully applied in image classification tasks recently. In this paper, we show that, unlike convolution neural networks (CNNs)that can be improved by stacking more convolutional layers, the performance of ViTs saturate fast when scaled to be deeper. More specifically, we empirically observe that such scaling difficulty is caused by the attention collapse issue: as the transformer goes deeper, the attention maps gradually become similar and even much the same after certain layers. In other words, the feature maps tend to be identical in the top layers of deep ViT models. This fact demonstrates that in deeper layers of ViTs, the self-attention mechanism fails to learn effective concepts for representation learning and hinders the model from getting expected performance gain. Based on above observation, we propose a simple yet effective method, named Re-attention, to re-generate the attention maps to increase their diversity at different layers with negligible computation and memory cost. The pro-posed method makes it feasible to train deeper ViT models with consistent performance improvements via minor modification to existing ViT models. Notably, when training a deep ViT model with 32 transformer blocks, the Top-1 classification accuracy can be improved by 1.6% on ImageNet. Code is publicly available at https://github.com/zhoudaquan/dvit_repo.

Qibin Hou, Daquan Zhou, Jiashi Feng

Recent studies on mobile network design have demonstrated the remarkable effectiveness of channel attention (e.g., the Squeeze-and-Excitation attention) for lifting model performance, but they generally neglect the positional information, which is important for generating spatially selective attention maps. In this paper, we propose a novel attention mechanism for mobile networks by embedding positional information into channel attention, which we call "coordinate attention". Unlike channel attention that transforms a feature tensor to a single feature vector via 2D global pooling, the coordinate attention factorizes channel attention into two 1D feature encoding processes that aggregate features along the two spatial directions, respectively. In this way, long-range dependencies can be captured along one spatial direction and meanwhile precise positional information can be preserved along the other spatial direction. The resulting feature maps are then encoded separately into a pair of direction-aware and position-sensitive attention maps that can be complementarily applied to the input feature map to augment the representations of the objects of interest. Our coordinate attention is simple and can be flexibly plugged into classic mobile networks, such as MobileNetV2, MobileNeXt, and EfficientNet with nearly no computational overhead. Extensive experiments demonstrate that our coordinate attention is not only beneficial to ImageNet classification but more interestingly, behaves better in down-stream tasks, such as object detection and semantic segmentation. Code is available at https://github.com/Andrew-Qibin/CoordAttention.

Weimin Wang, Jiawei Liu, Zhijie Lin et al.

The growing demand for high-fidelity video generation from textual descriptions has catalyzed significant research in this field. In this work, we introduce MagicVideo-V2 that integrates the text-to-image model, video motion generator, reference image embedding module and frame interpolation module into an end-to-end video generation pipeline. Benefiting from these architecture designs, MagicVideo-V2 can generate an aesthetically pleasing, high-resolution video with remarkable fidelity and smoothness. It demonstrates superior performance over leading Text-to-Video systems such as Runway, Pika 1.0, Morph, Moon Valley and Stable Video Diffusion model via user evaluation at large scale.

Xuanyi Li, Daquan Zhou, Chenxu Zhang et al.

The recently developed Sora model [1] has exhibited remarkable capabilities in video generation, sparking intense discussions regarding its ability to simulate real-world phenomena. Despite its growing popularity, there is a lack of established metrics to evaluate its fidelity to real-world physics quantitatively. In this paper, we introduce a new benchmark that assesses the quality of the generated videos based on their adherence to real-world physics principles. We employ a method that transforms the generated videos into 3D models, leveraging the premise that the accuracy of 3D reconstruction is heavily contingent on the video quality. From the perspective of 3D reconstruction, we use the fidelity of the geometric constraints satisfied by the constructed 3D models as a proxy to gauge the extent to which the generated videos conform to real-world physics rules. Project page: https://sora-geometrical-consistency.github.io/

Tianwei Xiong, Yuqing Wang, Daquan Zhou et al.

The efficacy of video generation models heavily depends on the quality of their training datasets. Most previous video generation models are trained on short video clips, while recently there has been increasing interest in training long video generation models directly on longer videos. However, the lack of such high-quality long videos impedes the advancement of long video generation. To promote research in long video generation, we desire a new dataset with four key features essential for training long video generation models: (1) long videos covering at least 10 seconds, (2) long-take videos without cuts, (3) large motion and diverse contents, and (4) temporally dense captions. To achieve this, we introduce a new pipeline for selecting high-quality long-take videos and generating temporally dense captions. Specifically, we define a set of metrics to quantitatively assess video quality including scene cuts, dynamic degrees, and semantic-level quality, enabling us to filter high-quality long-take videos from a large amount of source videos. Subsequently, we develop a hierarchical video captioning pipeline to annotate long videos with temporally-dense captions. With this pipeline, we curate the first long-take video dataset, LVD-2M, comprising 2 million long-take videos, each covering more than 10 seconds and annotated with temporally dense captions. We further validate the effectiveness of LVD-2M by fine-tuning video generation models to generate long videos with dynamic motions. We believe our work will significantly contribute to future research in long video generation.

Yupeng Zhou, Lianghua Huang, Zhifan Wu et al.

In this work, we propose Mutual Forcing, a framework for fast autoregressive audio-video generation with long-horizon audio-video synchronization. Our approach addresses two key challenges: joint audio-video modeling and fast autoregressive generation. To ease joint audio-video optimization, we adopt a two-stage training strategy: we first train uni-modal generators and then couple them into a unified audio-video model for joint training on paired data. For streaming generation, we ask whether a native fast causal audio-video model can be trained directly, instead of following existing streaming distillation pipelines that typically train a bidirectional model first and then convert it into a causal generator through multiple distillation stages. Our answer is Mutual Forcing, which builds directly on native autoregressive model and integrates few-step and multi-step generation within a single weight-shared model, enabling self-distillation and improved training-inference consistency. The multi-step mode improves the few-step mode via self-distillation, while the few-step mode generates historical context during training to improve training-inference consistency; because the two modes share parameters, these two effects reinforce each other within a single model. Compared with prior approaches such as Self-Forcing, Mutual Forcing removes the need for an additional bidirectional teacher model, supports more flexible training sequence lengths, reduces training overhead, and allows the model to improve directly from real paired data rather than a fixed teacher. Experiments show that Mutual Forcing matches or surpasses strong baselines that require around 50 sampling steps while using only 4 to 8 steps, demonstrating substantial advantages in both efficiency and quality. The project page is available at https://mutualforcing.github.io.

Xuying Zhang, Yupeng Zhou, Kai Wang et al.

Novel view synthesis (NVS) is a cornerstone for image-to-3d creation. However, existing works still struggle to maintain consistency between the generated views and the input views, especially when there is a significant camera pose difference, leading to poor-quality 3D geometries and textures. We attribute this issue to their treatment of all target views with equal priority according to our empirical observation that the target views closer to the input views exhibit higher fidelity. With this inspiration, we propose AR-1-to-3, a novel next-view prediction paradigm based on diffusion models that first generates views close to the input views, which are then utilized as contextual information to progressively synthesize farther views. To encode the generated view subsequences as local and global conditions for the next-view prediction, we accordingly develop a stacked local feature encoding strategy (Stacked-LE) and an LSTM-based global feature encoding strategy (LSTM-GE). Extensive experiments demonstrate that our method significantly improves the consistency between the generated views and the input views, producing high-fidelity 3D assets.

Junsong Chen, Yuyang Zhao, Jincheng Yu et al.

We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation.

Hongwei Yi, Tian Ye, Shitong Shao et al.

We present MagicInfinite, a novel diffusion Transformer (DiT) framework that overcomes traditional portrait animation limitations, delivering high-fidelity results across diverse character types-realistic humans, full-body figures, and stylized anime characters. It supports varied facial poses, including back-facing views, and animates single or multiple characters with input masks for precise speaker designation in multi-character scenes. Our approach tackles key challenges with three innovations: (1) 3D full-attention mechanisms with a sliding window denoising strategy, enabling infinite video generation with temporal coherence and visual quality across diverse character styles; (2) a two-stage curriculum learning scheme, integrating audio for lip sync, text for expressive dynamics, and reference images for identity preservation, enabling flexible multi-modal control over long sequences; and (3) region-specific masks with adaptive loss functions to balance global textual control and local audio guidance, supporting speaker-specific animations. Efficiency is enhanced via our innovative unified step and cfg distillation techniques, achieving a 20x inference speed boost over the basemodel: generating a 10 second 540x540p video in 10 seconds or 720x720p in 30 seconds on 8 H100 GPUs, without quality loss. Evaluations on our new benchmark demonstrate MagicInfinite's superiority in audio-lip synchronization, identity preservation, and motion naturalness across diverse scenarios. It is publicly available at https://www.hedra.com/, with examples at https://magicinfinite.github.io/.

Rui Ma, Qiang Zhou, Yizhu Jin et al.

Copyright law confers upon creators the exclusive rights to reproduce, distribute, and monetize their creative works. However, recent progress in text-to-image generation has introduced formidable challenges to copyright enforcement. These technologies enable the unauthorized learning and replication of copyrighted content, artistic creations, and likenesses, leading to the proliferation of unregulated content. Notably, models like stable diffusion, which excel in text-to-image synthesis, heighten the risk of copyright infringement and unauthorized distribution.Machine unlearning, which seeks to eradicate the influence of specific data or concepts from machine learning models, emerges as a promising solution by eliminating the \enquote{copyright memories} ingrained in diffusion models. Yet, the absence of comprehensive large-scale datasets and standardized benchmarks for evaluating the efficacy of unlearning techniques in the copyright protection scenarios impedes the development of more effective unlearning methods. To address this gap, we introduce a novel pipeline that harmonizes CLIP, ChatGPT, and diffusion models to curate a dataset. This dataset encompasses anchor images, associated prompts, and images synthesized by text-to-image models. Additionally, we have developed a mixed metric based on semantic and style information, validated through both human and artist assessments, to gauge the effectiveness of unlearning approaches. Our dataset, benchmark library, and evaluation metrics will be made publicly available to foster future research and practical applications (https://rmpku.github.io/CPDM-page/, website / http://149.104.22.83/unlearning.tar.gz, dataset).

Lin Xu, Ningxin Peng, Daquan Zhou et al.

Dialogue state tracking (DST) aims to record user queries and goals during a conversational interaction achieved by maintaining a predefined set of slots and their corresponding values. Current approaches decide slot values opaquely, while humans usually adopt a more deliberate approach by collecting information from relevant dialogue turns and then reasoning the appropriate values. In this work, we focus on the steps needed to figure out slot values by proposing a model named Chain-of-Thought-Explanation (CoTE) for the DST task. CoTE, which is built on the generative DST framework, is designed to create detailed explanations step by step after determining the slot values. This process leads to more accurate and reliable slot values. More-over, to improve the reasoning ability of the CoTE, we further construct more fluent and high-quality explanations with automatic paraphrasing, leading the method CoTE-refined. Experimental results on three widely recognized DST benchmarks-MultiWOZ 2.2, WoZ 2.0, and M2M-demonstrate the remarkable effectiveness of the CoTE. Furthermore, through a meticulous fine-grained analysis, we observe significant benefits of our CoTE on samples characterized by longer dialogue turns, user responses, and reasoning steps.

Yupeng Zhou, Zhen Li, Ziheng Ouyang et al.

Encoding videos into discrete tokens could align with text tokens to facilitate concise and unified multi-modal LLMs, yet introducing significant spatiotemporal compression compared to continuous video representation. Previous discrete video VAEs experienced unstable training, long training time, and degraded reconstruction quality. Given the easier training and superior performance of continuous VAEs, an intuitive idea is to enhance discrete video VAEs by leveraging continuous VAEs. After rethinking the intrinsic link between discrete and continuous representations, we found that FSQ could effectively preserve pre-trained continuous VAE priors compared to other quantization methods. By leveraging continuous VAE priors, it converges several times faster than training from scratch and achieves superior performance at convergence. Meanwhile, two structural improvements are proposed. First, inspired by how continuous VAEs enhance reconstruction via enlarged latent dimensions, we introduce a multi-token quantization mechanism, which achieves nearly a 1 dB improvement in PSNR without compromising the token compression ratio. Second, to tackle reconstruction challenges in high-compression video VAEs, we strengthen first-frame reconstruction, enabling the causal VAE to leverage this information in subsequent frames and markedly improving the performance of 4 x 16 x 16 discrete VAEs. Furthermore, we propose a joint discrete-continuous optimization scheme that unifies the two paradigms and, for the first time, achieves competitive performance on both continuous and discrete representations within a single network. We name our method OneVAE to reflect this connection.