Zhiyu Tan

Jiacheng Zhang, Yang Jiao, Shaoxiang Chen et al.

Recently, Multimodal Large Language Models (MLLMs) have made significant progress in the video comprehension field. Despite remarkable content reasoning and instruction following capabilities they demonstrated, the hallucination problem of these VideoLLMs is less explored compared with its counterpart in the image domain. To mitigate this gap, we propose EventHallusion, a novel benchmark that focuses on assessing the VideoLLMs' hallucination toward event, the crux of video analysis. From a hallucination attribution perspective, our EventHallusion benchmark is curated to assess a VideoLLM's susceptibility toward language priors and vision-language biases. On the other hand, we also propose a simple yet effective method, called Temporal Contrastive Decoding (TCD), to tackle the hallucination problems of VideoLLMs. The proposed TCD method rectifies the model's bias toward its priors during the decoding stage by comparing the original video with a modified version, in which temporal cues are disrupted. Through comprehensive evaluation of eight open-source and two closed-source VideoLLMs on the proposed EventHallusion benchmark, we observe that the open-source models suffer significantly from hallucination problems, whereas the closed-source ones perform markedly better. By further equipping open-source VideoLLMs with the proposed TCD approach, evident performance improvements are achieved across most metrics in the EventHallusion benchmark. Our codes and benchmark data are available at https://github.com/Stevetich/EventHallusion.

Zhiyu Tan, Xiaomeng Yang, Luozheng Qin et al.

The quality of video-text pairs fundamentally determines the upper bound of text-to-video models. Currently, the datasets used for training these models suffer from significant shortcomings, including low temporal consistency, poor-quality captions, substandard video quality, and imbalanced data distribution. The prevailing video curation process, which depends on image models for tagging and manual rule-based curation, leads to a high computational load and leaves behind unclean data. As a result, there is a lack of appropriate training datasets for text-to-video models. To address this problem, we present VidGen-1M, a superior training dataset for text-to-video models. Produced through a coarse-to-fine curation strategy, this dataset guarantees high-quality videos and detailed captions with excellent temporal consistency. When used to train the video generation model, this dataset has led to experimental results that surpass those obtained with other models.

Ziyi Gao, Kai Chen, Zhipeng Wei et al.

Recent diffusion-based unrestricted attacks generate imperceptible adversarial examples with high transferability compared to previous unrestricted attacks and restricted attacks. However, existing works on diffusion-based unrestricted attacks are mostly focused on images yet are seldom explored in videos. In this paper, we propose the Recursive Token Merging for Video Diffusion-based Unrestricted Adversarial Attack (ReToMe-VA), which is the first framework to generate imperceptible adversarial video clips with higher transferability. Specifically, to achieve spatial imperceptibility, ReToMe-VA adopts a Timestep-wise Adversarial Latent Optimization (TALO) strategy that optimizes perturbations in diffusion models' latent space at each denoising step. TALO offers iterative and accurate updates to generate more powerful adversarial frames. TALO can further reduce memory consumption in gradient computation. Moreover, to achieve temporal imperceptibility, ReToMe-VA introduces a Recursive Token Merging (ReToMe) mechanism by matching and merging tokens across video frames in the self-attention module, resulting in temporally consistent adversarial videos. ReToMe concurrently facilitates inter-frame interactions into the attack process, inducing more diverse and robust gradients, thus leading to better adversarial transferability. Extensive experiments demonstrate the efficacy of ReToMe-VA, particularly in surpassing state-of-the-art attacks in adversarial transferability by more than 14.16% on average.

Hao Yang, Zhiyu Tan, Jia Gong et al.

We present Omni-Video 2, a scalable and computationally efficient model that connects pretrained multimodal large-language models (MLLMs) with video diffusion models for unified video generation and editing. Our key idea is to exploit the understanding and reasoning capabilities of MLLMs to produce explicit target captions to interpret user instructions. In this way, the rich contextual representations from the understanding model are directly used to guide the generative process, thereby improving performance on complex and compositional editing. Moreover, a lightweight adapter is developed to inject multimodal conditional tokens into pretrained text-to-video diffusion models, allowing maximum reuse of their powerful generative priors in a parameter-efficient manner. Benefiting from these designs, we scale up Omni-Video 2 to a 14B video diffusion model on meticulously curated training data with quality, supporting high quality text-to-video generation and various video editing tasks such as object removal, addition, background change, complex motion editing, \emph{etc.} We evaluate the performance of Omni-Video 2 on the FiVE benchmark for fine-grained video editing and the VBench benchmark for text-to-video generation. The results demonstrate its superior ability to follow complex compositional instructions in video editing, while also achieving competitive or superior quality in video generation tasks.

Mengping Yang, Zhiyu Tan, Binglei Li et al.

Recent breakthroughs in Diffusion Transformers (DiTs) have revolutionized the field of visual synthesis due to their superior scalability. To facilitate DiTs' capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment. However, the underlying mechanisms governing representation learning within DiTs are not well understood. To this end, we first systematically investigate the representation dynamics of DiTs. Through analyzing the evolution and influence of internal representations under various settings, we reveal that representation diversity across blocks is a crucial factor for effective learning. Based on this key insight, we propose DiverseDiT, a novel framework that explicitly promotes representation diversity. DiverseDiT incorporates long residual connections to diversify input representations across blocks and a representation diversity loss to encourage blocks to learn distinct features. Extensive experiments on ImageNet 256x256 and 512x512 demonstrate that our DiverseDiT yields consistent performance gains and convergence acceleration when applied to different backbones with various sizes, even when tested on the challenging one-step generation setting. Furthermore, we show that DiverseDiT is complementary to existing representation learning techniques, leading to further performance gains. Our work provides valuable insights into the representation learning dynamics of DiTs and offers a practical approach for enhancing their performance.

Binglei Li, Mengping Yang, Zhiyu Tan et al.

Recent breakthroughs of transformer-based diffusion models, particularly with Multimodal Diffusion Transformers (MMDiT) driven models like FLUX and Qwen Image, have facilitated thrilling experiences in text-to-image generation and editing. To understand the internal mechanism of MMDiT-based models, existing methods tried to analyze the effect of specific components like positional encoding and attention layers. Yet, a comprehensive understanding of how different blocks and their interactions with textual conditions contribute to the synthesis process remains elusive. In this paper, we first develop a systematic pipeline to comprehensively investigate each block's functionality by removing, disabling and enhancing textual hidden-states at corresponding blocks. Our analysis reveals that 1) semantic information appears in earlier blocks and finer details are rendered in later blocks, 2) removing specific blocks is usually less disruptive than disabling text conditions, and 3) enhancing textual conditions in selective blocks improves semantic attributes. Building on these observations, we further propose novel training-free strategies for improved text alignment, precise editing, and acceleration. Extensive experiments demonstrated that our method outperforms various baselines and remains flexible across text-to-image generation, image editing, and inference acceleration. Our method improves T2I-Combench++ from 56.92% to 63.00% and GenEval from 66.42% to 71.63% on SD3.5, without sacrificing synthesis quality. These results advance understanding of MMDiT models and provide valuable insights to unlock new possibilities for further improvements.

Zhiyu Tan, Zichao Dong, Cheng Zhang et al.

Semantic occupancy prediction aims to infer dense geometry and semantics of surroundings for an autonomous agent to operate safely in the 3D environment. Existing occupancy prediction methods are almost entirely trained on human-annotated volumetric data. Although of high quality, the generation of such 3D annotations is laborious and costly, restricting them to a few specific object categories in the training dataset. To address this limitation, this paper proposes Open Vocabulary Occupancy (OVO), a novel approach that allows semantic occupancy prediction of arbitrary classes but without the need for 3D annotations during training. Keys to our approach are (1) knowledge distillation from a pre-trained 2D open-vocabulary segmentation model to the 3D occupancy network, and (2) pixel-voxel filtering for high-quality training data generation. The resulting framework is simple, compact, and compatible with most state-of-the-art semantic occupancy prediction models. On NYUv2 and SemanticKITTI datasets, OVO achieves competitive performance compared to supervised semantic occupancy prediction approaches. Furthermore, we conduct extensive analyses and ablation studies to offer insights into the design of the proposed framework. Our code is publicly available at https://github.com/dzcgaara/OVO.

Yiqi Jiang, Zhiyu Tan, Junyan Wang et al.

In conventional object detection frameworks, a backbone body inherited from image recognition models extracts deep latent features and then a neck module fuses these latent features to capture information at different scales. As the resolution in object detection is much larger than in image recognition, the computational cost of the backbone often dominates the total inference cost. This heavy-backbone design paradigm is mostly due to the historical legacy when transferring image recognition models to object detection rather than an end-to-end optimized design for object detection. In this work, we show that such paradigm indeed leads to sub-optimal object detection models. To this end, we propose a novel heavy-neck paradigm, GiraffeDet, a giraffe-like network for efficient object detection. The GiraffeDet uses an extremely lightweight backbone and a very deep and large neck module which encourages dense information exchange among different spatial scales as well as different levels of latent semantics simultaneously. This design paradigm allows detectors to process the high-level semantic information and low-level spatial information at the same priority even in the early stage of the network, making it more effective in detection tasks. Numerical evaluations on multiple popular object detection benchmarks show that GiraffeDet consistently outperforms previous SOTA models across a wide spectrum of resource constraints. The source code is available at https://github.com/jyqi/GiraffeDet.

Zhenhong Sun, Ming Lin, Xiuyu Sun et al.

In object detection, the detection backbone consumes more than half of the overall inference cost. Recent researches attempt to reduce this cost by optimizing the backbone architecture with the help of Neural Architecture Search (NAS). However, existing NAS methods for object detection require hundreds to thousands of GPU hours of searching, making them impractical in fast-paced research and development. In this work, we propose a novel zero-shot NAS method to address this issue. The proposed method, named MAE-DET, automatically designs efficient detection backbones via the Maximum Entropy Principle without training network parameters, reducing the architecture design cost to nearly zero yet delivering the state-of-the-art (SOTA) performance. Under the hood, MAE-DET maximizes the differential entropy of detection backbones, leading to a better feature extractor for object detection under the same computational budgets. After merely one GPU day of fully automatic design, MAE-DET innovates SOTA detection backbones on multiple detection benchmark datasets with little human intervention. Comparing to ResNet-50 backbone, MAE-DET is $+2.0\%$ better in mAP when using the same amount of FLOPs/parameters, and is $1.54$ times faster on NVIDIA V100 at the same mAP. Code and pre-trained models are available at https://github.com/alibaba/lightweight-neuralarchitecture-search.

Yibin Wang, Zhiyu Tan, Junyan Wang et al.

Recent advances in text-to-video (T2V) generative models have shown impressive capabilities. However, these models are still inadequate in aligning synthesized videos with human preferences (e.g., accurately reflecting text descriptions), which is particularly difficult to address, as human preferences are subjective and challenging to formalize as objective functions. Existing studies train video quality assessment models that rely on human-annotated ratings for video evaluation but overlook the reasoning behind evaluations, limiting their ability to capture nuanced human criteria. Moreover, aligning T2V model using video-based human feedback remains unexplored. Therefore, this paper proposes LiFT, the first method designed to leverage human feedback for T2V model alignment. Specifically, we first construct a Human Rating Annotation dataset, LiFT-HRA, consisting of approximately 10k human annotations, each including a score and its corresponding rationale. Based on this, we train a reward model LiFT-Critic to learn reward function effectively, which serves as a proxy for human judgment, measuring the alignment between given videos and human expectations. Lastly, we leverage the learned reward function to align the T2V model by maximizing the reward-weighted likelihood. As a case study, we apply our pipeline to CogVideoX-2B, showing that the fine-tuned model outperforms the CogVideoX-5B across all 16 metrics, highlighting the potential of human feedback in improving the alignment and quality of synthesized videos.

Junyan Wang, Zhenhong Sun, Zhiyu Tan et al.

Vanilla text-to-image diffusion models struggle with generating accurate human images, commonly resulting in imperfect anatomies such as unnatural postures or disproportionate limbs.Existing methods address this issue mostly by fine-tuning the model with extra images or adding additional controls -- human-centric priors such as pose or depth maps -- during the image generation phase. This paper explores the integration of these human-centric priors directly into the model fine-tuning stage, essentially eliminating the need for extra conditions at the inference stage. We realize this idea by proposing a human-centric alignment loss to strengthen human-related information from the textual prompts within the cross-attention maps. To ensure semantic detail richness and human structural accuracy during fine-tuning, we introduce scale-aware and step-wise constraints within the diffusion process, according to an in-depth analysis of the cross-attention layer. Extensive experiments show that our method largely improves over state-of-the-art text-to-image models to synthesize high-quality human images based on user-written prompts. Project page: \url{https://hcplayercvpr2024.github.io}.

Xiaomeng Yang, Zhiyu Tan, Hao Li

Video foundation models have achieved significant advancement with the help of network upgrade as well as model scale-up. However, they are still hard to meet requirements of applications due to unsatisfied generation quality. To solve this problem, we propose to align video foundation models with human preferences from the perspective of post-training in this paper. Consequently, we introduce an Iterative Preference Optimization strategy to enhance generated video quality by incorporating human feedback. Specifically, IPO exploits a critic model to justify video generations for pairwise ranking as in Direct Preference Optimization or point-wise scoring as in Kahneman-Tversky Optimization. Given this, IPO optimizes video foundation models with guidance of signals from preference feedback, which helps improve generated video quality in subject consistency, motion smoothness and aesthetic quality, etc. In addition, IPO incorporates the critic model with the multi-modality large language model, which enables it to automatically assign preference labels without need of retraining or relabeling. In this way, IPO can efficiently perform multi-round preference optimization in an iterative manner, without the need of tediously manual labeling. Comprehensive experiments demonstrate that the proposed IPO can effectively improve the video generation quality of a pretrained model and help a model with only 2B parameters surpass the one with 5B parameters. Besides, IPO achieves new state-of-the-art performance on VBench benchmark.

Zhiyu Tan, Mengping Yang, Luozheng Qin et al.

One critical prerequisite for faithful text-to-image generation is the accurate understanding of text inputs. Existing methods leverage the text encoder of the CLIP model to represent input prompts. However, the pre-trained CLIP model can merely encode English with a maximum token length of 77. Moreover, the model capacity of the text encoder from CLIP is relatively limited compared to Large Language Models (LLMs), which offer multilingual input, accommodate longer context, and achieve superior text representation. In this paper, we investigate LLMs as the text encoder to improve the language understanding in text-to-image generation. Unfortunately, training text-to-image generative model with LLMs from scratch demands significant computational resources and data. To this end, we introduce a three-stage training pipeline that effectively and efficiently integrates the existing text-to-image model with LLMs. Specifically, we propose a lightweight adapter that enables fast training of the text-to-image model using the textual representations from LLMs. Extensive experiments demonstrate that our model supports not only multilingual but also longer input context with superior image generation quality.

Luozheng Qin, Jia Gong, Yuqing Sun et al.

Chain-of-Thought (CoT) reasoning has been widely adopted to enhance Large Language Models (LLMs) by decomposing complex tasks into simpler, sequential subtasks. However, extending CoT to vision-language reasoning tasks remains challenging, as it often requires interpreting transitions of visual states to support reasoning. Existing methods often struggle with this due to limited capacity of modeling visual state transitions or incoherent visual trajectories caused by fragmented architectures. To overcome these limitations, we propose Uni-CoT, a Unified Chain-of-Thought framework that enables coherent and grounded multimodal reasoning within a single unified model. The key idea is to leverage a model capable of both image understanding and generation to reason over visual content and model evolving visual states. However, empowering a unified model to achieve that is non-trivial, given the high computational cost and the burden of training. To address this, Uni-CoT introduces a novel two-level reasoning paradigm: A Macro-Level CoT for high-level task planning and A Micro-Level CoT for subtask execution. This design significantly reduces the computational overhead. Furthermore, we introduce a structured training paradigm that combines interleaved image-text supervision for macro-level CoT with multi-task objectives for micro-level CoT. Together, these innovations allow Uni-CoT to perform scalable and coherent multi-modal reasoning. Furthermore, thanks to our design, all experiments can be efficiently completed using only 8 A100 GPUs with 80GB VRAM each. Experimental results on reasoning-driven image generation benchmark (WISE) and editing benchmarks (RISE and KRIS) indicates that Uni-CoT demonstrates SOTA performance and strong generalization, establishing Uni-CoT as a promising solution for multi-modal reasoning. Project Page and Code: https://sais-fuxi.github.io/projects/uni-cot/

Zhiyu Tan, Hao Yang, Luozheng Qin et al.

Notable breakthroughs in unified understanding and generation modeling have led to remarkable advancements in image understanding, reasoning, production and editing, yet current foundational models predominantly focus on processing images, creating a gap in the development of unified models for video understanding and generation. This report presents Omni-Video, an efficient and effective unified framework for video understanding, generation, as well as instruction-based editing. Our key insight is to teach existing multimodal large language models (MLLMs) to produce continuous visual clues that are used as the input of diffusion decoders, which produce high-quality videos conditioned on these visual clues. To fully unlock the potential of our system for unified video modeling, we integrate several technical improvements: 1) a lightweight architectural design that respectively attaches a vision head on the top of MLLMs and a adapter before the input of diffusion decoders, the former produce visual tokens for the latter, which adapts these visual tokens to the conditional space of diffusion decoders; and 2) an efficient multi-stage training scheme that facilitates a fast connection between MLLMs and diffusion decoders with limited data and computational resources. We empirically demonstrate that our model exhibits satisfactory generalization abilities across video generation, editing and understanding tasks.

Wenxu Qian, Chaoyue Wang, Hou Peng et al.

Video generation techniques have achieved remarkable advancements in visual quality, yet faithfully reproducing real-world physics remains elusive. Preference-based model post-training may improve physical consistency, but requires costly human-annotated datasets or reward models that are not yet feasible. To address these challenges, we present Real Data Preference Optimisation (RDPO), an annotation-free framework that distills physical priors directly from real-world videos. Specifically, the proposed RDPO reverse-samples real video sequences with a pre-trained generator to automatically build preference pairs that are statistically distinguishable in terms of physical correctness. A multi-stage iterative training schedule then guides the generator to obey physical laws increasingly well. Benefiting from the dynamic information explored from real videos, our proposed RDPO significantly improves the action coherence and physical realism of the generated videos. Evaluations on multiple benchmarks and human evaluations have demonstrated that RDPO achieves improvements across multiple dimensions. The source code and demonstration of this paper are available at: https://wwenxu.github.io/RDPO/

Hesen Chen, Junyan Wang, Zhiyu Tan et al.

Modern diffusion models encounter a fundamental trade-off between training efficiency and generation quality. While existing representation alignment methods, such as REPA, accelerate convergence through patch-wise alignment, they often fail to capture structural relationships within visual representations and ensure global distribution consistency between pretrained encoders and denoising networks. To address these limitations, we introduce SARA, a hierarchical alignment framework that enforces multi-level representation constraints: (1) patch-wise alignment to preserve local semantic details, (2) autocorrelation matrix alignment to maintain structural consistency within representations, and (3) adversarial distribution alignment to mitigate global representation discrepancies. Unlike previous approaches, SARA explicitly models both intra-representation correlations via self-similarity matrices and inter-distribution coherence via adversarial alignment, enabling comprehensive alignment across local and global scales. Experiments on ImageNet-256 show that SARA achieves an FID of 1.36 while converging twice as fast as REPA, surpassing recent state-of-the-art image generation methods. This work establishes a systematic paradigm for optimizing diffusion training through hierarchical representation alignment.

Zhiyu Tan, Junyan Wang, Hao Yang et al.

Text-to-video generation has demonstrated promising progress with the advent of diffusion models, yet existing approaches are limited by dataset quality and computational resources. To address these limitations, this paper presents a comprehensive approach that advances both data curation and model design. We introduce CFC-VIDS-1M, a high-quality video dataset constructed through a systematic coarse-to-fine curation pipeline. The pipeline first evaluates video quality across multiple dimensions, followed by a fine-grained stage that leverages vision-language models to enhance text-video alignment and semantic richness. Building upon the curated dataset's emphasis on visual quality and temporal coherence, we develop RACCOON, a transformer-based architecture with decoupled spatial-temporal attention mechanisms. The model is trained through a progressive four-stage strategy designed to efficiently handle the complexities of video generation. Extensive experiments demonstrate that our integrated approach of high-quality data curation and efficient training strategy generates visually appealing and temporally coherent videos while maintaining computational efficiency. We will release our dataset, code, and models.

Luozheng Qin, Jia Gong, Qian Qiao et al.

Unified multimodal models integrating visual understanding and generation face a fundamental challenge: visual generation incurs substantially higher computational costs than understanding, particularly for video. This imbalance motivates us to invert the conventional paradigm: rather than extending understanding-centric MLLMs to support generation, we propose Uni-ViGU, a framework that unifies video generation and understanding by extending a video generator as the foundation. We introduce a unified flow method that performs continuous flow matching for video and discrete flow matching for text within a single process, enabling coherent multimodal generation. We further propose a modality-driven MoE-based framework that augments Transformer blocks with lightweight layers for text generation while preserving generative priors. To repurpose generation knowledge for understanding, we design a bidirectional training mechanism with two stages: Knowledge Recall reconstructs input prompts to leverage learned text-video correspondences, while Capability Refinement fine-tunes on detailed captions to establish discriminative shared representations. Experiments demonstrate that Uni-ViGU achieves competitive performance on both video generation and understanding, validating generation-centric architectures as a scalable path toward unified multimodal intelligence. Project Page and Code: https://fr0zencrane.github.io/uni-vigu-page/.

Luozheng Qin, Zhiyu Tan, Mengping Yang et al.

Video Detailed Captioning (VDC) is a crucial task for vision-language bridging, enabling fine-grained descriptions of complex video content. In this paper, we first comprehensively benchmark current state-of-the-art approaches and systematically identified two critical limitations: biased capability towards specific captioning aspect and misalignment with human preferences. To address these deficiencies, we propose Cockatiel, a novel three-stage training pipeline that ensembles synthetic and human-aligned training for improving VDC performance. In the first stage, we derive a scorer from a meticulously annotated dataset to select synthetic captions high-performing on certain fine-grained video-caption alignment and human-preferred while disregarding others. Then, we train Cockatiel-13B, using this curated dataset to infuse it with assembled model strengths and human preferences. Finally, we further distill Cockatiel-8B from Cockatiel-13B for the ease of usage. Extensive quantitative and qualitative experiments reflect the effectiveness of our method, as we not only set new state-of-the-art performance on VDCSCORE in a dimension-balanced way but also surpass leading alternatives on human preference by a large margin as depicted by the human evaluation results.

Xiaomeng Yang, Zhiyu Tan, Xiaohui Zhong et al.

Scientific discovery increasingly relies on integrating heterogeneous, high-dimensional data across disciplines nowadays. While AI models have achieved notable success across various scientific domains, they typically remain domain-specific or lack the capability of simultaneously understanding and generating multimodal scientific data, particularly for high-dimensional data. Yet, many pressing global challenges and scientific problems are inherently cross-disciplinary and require coordinated progress across multiple fields. Here, we present FuXi-Uni, a native unified multimodal model for scientific understanding and high-fidelity generation across scientific domains within a single architecture. Specifically, FuXi-Uni aligns cross-disciplinary scientific tokens within natural language tokens and employs science decoder to reconstruct scientific tokens, thereby supporting both natural language conversation and scientific numerical prediction. Empirically, we validate FuXi-Uni in Earth science and Biomedicine. In Earth system modeling, the model supports global weather forecasting, tropical cyclone (TC) forecast editing, and spatial downscaling driven by only language instructions. FuXi-Uni generates 10-day global forecasts at 0.25° resolution that outperform the SOTA physical forecasting system. It shows superior performance for both TC track and intensity prediction relative to the SOTA physical model, and generates high-resolution regional weather fields that surpass standard interpolation baselines. Regarding biomedicine, FuXi-Uni outperforms leading multimodal large language models on multiple biomedical visual question answering benchmarks. By unifying heterogeneous scientific modalities within a native shared latent space while maintaining strong domain-specific performance, FuXi-Uni provides a step forward more general-purpose, multimodal scientific models.

Xiaomeng Yang, Zhiyu Tan, Junyan Wang et al.

Preference learning has become a central technique for aligning generative models with human expectations. Recently, it has been extended to diffusion models through methods like Direct Preference Optimization (DPO). However, existing approaches such as Diffusion-DPO suffer from two key challenges: timestep-dependent instability, caused by a mismatch between the reverse and forward diffusion processes and by high gradient variance in early noisy timesteps, and off-policy bias arising from the mismatch between optimization and data collection policies. We begin by analyzing the reverse diffusion trajectory and observe that instability primarily occurs at early timesteps with low importance weights. To address these issues, we first propose DPO-C\&M, a practical strategy that improves stability by clipping and masking uninformative timesteps while partially mitigating off-policy bias. Building on this, we introduce SDPO (Importance-Sampled Direct Preference Optimization), a principled framework that incorporates importance sampling into the objective to fully correct for off-policy bias and emphasize informative updates during the diffusion process. Experiments on CogVideoX-2B, CogVideoX-5B, and Wan2.1-1.3B demonstrate that both methods outperform standard Diffusion-DPO, with SDPO achieving superior VBench scores, human preference alignment, and training robustness. These results highlight the importance of timestep-aware, distribution-corrected optimization in diffusion-based preference learning.

Zhiyu Tan, WenXu Qian, Hesen Chen et al.

Diffusion models have established themselves as the de facto primary paradigm in visual generative modeling, revolutionizing the field through remarkable success across various diverse applications ranging from high-quality image synthesis to temporal aware video generation. Despite these advancements, three fundamental limitations persist, including 1) discrepancy between training and inference processes, 2) progressive information leakage throughout the noise corruption procedures, and 3) inherent constraints preventing effective integration of modern optimization criteria like perceptual and adversarial loss. To mitigate these critical challenges, we in this paper present a novel end-to-end learning paradigm that establishes direct optimization from the final generated samples to initial noises. Our proposed End-to-End Differentiable Diffusion, dubbed E2ED^2, introduces several key improvements: it eliminates the sequential training-sampling mismatch and intermediate information leakage via conceptualizing training as a direct transformation from isotropic Gaussian noise to the target data distribution. Additionally, such training framework enables seamless incorporation of adversarial and perceptual losses into the core optimization objective. Comprehensive evaluation across standard benchmarks including COCO30K and HW30K reveals that our method achieves substantial performance gains in terms of Fréchet Inception Distance (FID) and CLIP score, even with fewer sampling steps (less than 4). Our findings highlight that the end-to-end mechanism might pave the way for more robust and efficient solutions, \emph{i.e.,} combining diffusion stability with GAN-like discriminative optimization in an end-to-end manner.

Zhiyu Tan, Xiaomeng Yang, Luozheng Qin et al.

The recent advancements in text-to-image generative models have been remarkable. Yet, the field suffers from a lack of evaluation metrics that accurately reflect the performance of these models, particularly lacking fine-grained metrics that can guide the optimization of the models. In this paper, we propose EvalAlign, a metric characterized by its accuracy, stability, and fine granularity. Our approach leverages the capabilities of Multimodal Large Language Models (MLLMs) pre-trained on extensive data. We develop evaluation protocols that focus on two key dimensions: image faithfulness and text-image alignment. Each protocol comprises a set of detailed, fine-grained instructions linked to specific scoring options, enabling precise manual scoring of the generated images. We supervised fine-tune (SFT) the MLLM to align with human evaluative judgments, resulting in a robust evaluation model. Our evaluation across 24 text-to-image generation models demonstrate that EvalAlign not only provides superior metric stability but also aligns more closely with human preferences than existing metrics, confirming its effectiveness and utility in model assessment.

Yichen Qian, Ming Lin, Xiuyu Sun et al.

One critical component in lossy deep image compression is the entropy model, which predicts the probability distribution of the quantized latent representation in the encoding and decoding modules. Previous works build entropy models upon convolutional neural networks which are inefficient in capturing global dependencies. In this work, we propose a novel transformer-based entropy model, termed Entroformer, to capture long-range dependencies in probability distribution estimation effectively and efficiently. Different from vision transformers in image classification, the Entroformer is highly optimized for image compression, including a top-k self-attention and a diamond relative position encoding. Meanwhile, we further expand this architecture with a parallel bidirectional context model to speed up the decoding process. The experiments show that the Entroformer achieves state-of-the-art performance on image compression while being time-efficient.

Zhenhong Sun, Zhiyu Tan, Xiuyu Sun et al.

Compression standards have been used to reduce the cost of image storage and transmission for decades. In recent years, learned image compression methods have been proposed and achieved compelling performance to the traditional standards. However, in these methods, a set of different networks are used for various compression rates, resulting in a high cost in model storage and training. Although some variable-rate approaches have been proposed to reduce the cost by using a single network, most of them brought some performance degradation when applying fine rate control. To enable variable-rate control without sacrificing the performance, we propose an efficient Interpolation Variable-Rate (IVR) network, by introducing a handy Interpolation Channel Attention (InterpCA) module in the compression network. With the use of two hyperparameters for rate control and linear interpolation, the InterpCA achieves a fine PSNR interval of 0.001 dB and a fine rate interval of 0.0001 Bits-Per-Pixel (BPP) with 9000 rates in the IVR network. Experimental results demonstrate that the IVR network is the first variable-rate learned method that outperforms VTM 9.0 (intra) in PSNR and Multiscale Structural Similarity (MS-SSIM).

Zhenhong Sun, Zhiyu Tan, Xiuyu Sun et al.

The framework of dominant learned video compression methods is usually composed of motion prediction modules as well as motion vector and residual image compression modules, suffering from its complex structure and error propagation problem. Approaches have been proposed to reduce the complexity by replacing motion prediction modules with implicit flow networks. Error propagation aware training strategy is also proposed to alleviate incremental reconstruction errors from previously decoded frames. Although these methods have brought some improvement, little attention has been paid to the framework itself. Inspired by the success of learned image compression through simplifying the framework with a single deep neural network, it is natural to expect a better performance in video compression via a simple yet appropriate framework. Therefore, we propose a framework to directly compress raw-pixel frames (rather than residual images), where no extra motion prediction module is required. Instead, an entropy model is used to estimate the spatiotemporal redundancy in a latent space rather than pixel level, which significantly reduces the complexity of the framework. Specifically, the whole framework is a compression module, consisting of a unified auto-encoder which produces identically distributed latents for all frames, and a spatiotemporal entropy estimation model to minimize the entropy of these latents. Experiments showed that the proposed method outperforms state-of-the-art (SOTA) performance under the metric of multiscale structural similarity (MS-SSIM) and achieves competitive results under the metric of PSNR.

Yichen Qian, Zhiyu Tan, Xiuyu Sun et al.

In recent deep image compression neural networks, the entropy model plays a critical role in estimating the prior distribution of deep image encodings. Existing methods combine hyperprior with local context in the entropy estimation function. This greatly limits their performance due to the absence of a global vision. In this work, we propose a novel Global Reference Model for image compression to effectively leverage both the local and the global context information, leading to an enhanced compression rate. The proposed method scans decoded latents and then finds the most relevant latent to assist the distribution estimating of the current latent. A by-product of this work is the innovation of a mean-shifting GDN module that further improves the performance. Experimental results demonstrate that the proposed model outperforms the rate-distortion performance of most of the state-of-the-art methods in the industry.