Guoqiang Wei

Yan Zeng, Hanbo Zhang, Jiani Zheng et al. · bytedance

Recent advancements in Large Language Models (LLMs) such as GPT4 have displayed exceptional multi-modal capabilities in following open-ended instructions given images. However, the performance of these models heavily relies on design choices such as network structures, training data, and training strategies, and these choices have not been extensively discussed in the literature, making it difficult to quantify progress in this field. To address this issue, this paper presents a systematic and comprehensive study, quantitatively and qualitatively, on training such models. We implement over 20 variants with controlled settings. Concretely, for network structures, we compare different LLM backbones and model designs. For training data, we investigate the impact of data and sampling strategies. For instructions, we explore the influence of diversified prompts on the instruction-following ability of the trained models. For benchmarks, we contribute the first, to our best knowledge, comprehensive evaluation set including both image and video tasks through crowd-sourcing. Based on our findings, we present Lynx, which performs the most accurate multi-modal understanding while keeping the best multi-modal generation ability compared to existing open-sourced GPT4-style models.

Guoqiang Wei, Zhizheng Zhang, Cuiling Lan et al.

The three existing dominant network families, i.e., CNNs, Transformers, and MLPs, differ from each other mainly in the ways of fusing spatial contextual information, leaving designing more effective token-mixing mechanisms at the core of backbone architecture development. In this work, we propose an innovative token-mixer, dubbed Active Token Mixer (ATM), to actively incorporate flexible contextual information distributed across different channels from other tokens into the given query token. This fundamental operator actively predicts where to capture useful contexts and learns how to fuse the captured contexts with the query token at channel level. In this way, the spatial range of token-mixing can be expanded to a global scope with limited computational complexity, where the way of token-mixing is reformed. We take ATM as the primary operator and assemble ATMs into a cascade architecture, dubbed ATMNet. Extensive experiments demonstrate that ATMNet is generally applicable and comprehensively surpasses different families of SOTA vision backbones by a clear margin on a broad range of vision tasks, including visual recognition and dense prediction tasks. Code is available at https://github.com/microsoft/ActiveMLP.

Team Seedance, De Chen, Liyang Chen et al. · gatech

Seedance 2.0 is a new native multi-modal audio-video generation model, officially released in China in early February 2026. Compared with its predecessors, Seedance 1.0 and 1.5 Pro, Seedance 2.0 adopts a unified, highly efficient, and large-scale architecture for multi-modal audio-video joint generation. This allows it to support four input modalities: text, image, audio, and video, by integrating one of the most comprehensive suites of multi-modal content reference and editing capabilities available in the industry to date. It delivers substantial, well-rounded improvements across all key sub-dimensions of video and audio generation. In both expert evaluations and public user tests, the model has demonstrated performance on par with the leading levels in the field. Seedance 2.0 supports direct generation of audio-video content with durations ranging from 4 to 15 seconds, with native output resolutions of 480p and 720p. For multi-modal inputs as reference, its current open platform supports up to 3 video clips, 9 images, and 3 audio clips. In addition, we provide Seedance 2.0 Fast version, an accelerated variant of Seedance 2.0 designed to boost generation speed for low-latency scenarios. Seedance 2.0 has delivered significant improvements to its foundational generation capabilities and multi-modal generation performance, bringing an enhanced creative experience for end users.

Xin Li, Cuiling Lan, Guoqiang Wei et al.

Vision transformer has demonstrated great potential in abundant vision tasks. However, it also inevitably suffers from poor generalization capability when the distribution shift occurs in testing (i.e., out-of-distribution data). To mitigate this issue, we propose a novel method, Semantic-aware Message Broadcasting (SAMB), which enables more informative and flexible feature alignment for unsupervised domain adaptation (UDA). Particularly, we study the attention module in the vision transformer and notice that the alignment space using one global class token lacks enough flexibility, where it interacts information with all image tokens in the same manner but ignores the rich semantics of different regions. In this paper, we aim to improve the richness of the alignment features by enabling semantic-aware adaptive message broadcasting. Particularly, we introduce a group of learned group tokens as nodes to aggregate the global information from all image tokens, but encourage different group tokens to adaptively focus on the message broadcasting to different semantic regions. In this way, our message broadcasting encourages the group tokens to learn more informative and diverse information for effective domain alignment. Moreover, we systematically study the effects of adversarial-based feature alignment (ADA) and pseudo-label based self-training (PST) on UDA. We find that one simple two-stage training strategy with the cooperation of ADA and PST can further improve the adaptation capability of the vision transformer. Extensive experiments on DomainNet, OfficeHome, and VisDA-2017 demonstrate the effectiveness of our methods for UDA.

Ganlin Yang, Guoqiang Wei, Zhizheng Zhang et al.

Most Neural Radiance Fields (NeRFs) exhibit limited generalization capabilities, which restrict their applicability in representing multiple scenes using a single model. To address this problem, existing generalizable NeRF methods simply condition the model on image features. These methods still struggle to learn precise global representations over diverse scenes since they lack an effective mechanism for interacting among different points and views. In this work, we unveil that 3D implicit representation learning can be significantly improved by mask-based modeling. Specifically, we propose masked ray and view modeling for generalizable NeRF (MRVM-NeRF), which is a self-supervised pretraining target to predict complete scene representations from partially masked features along each ray. With this pretraining target, MRVM-NeRF enables better use of correlations across different points and views as the geometry priors, which thereby strengthens the capability of capturing intricate details within the scenes and boosts the generalization capability across different scenes. Extensive experiments demonstrate the effectiveness of our proposed MRVM-NeRF on both synthetic and real-world datasets, qualitatively and quantitatively. Besides, we also conduct experiments to show the compatibility of our proposed method with various backbones and its superiority under few-shot cases.

Yan Zeng, Guoqiang Wei, Jiani Zheng et al.

Creating high-dynamic videos such as motion-rich actions and sophisticated visual effects poses a significant challenge in the field of artificial intelligence. Unfortunately, current state-of-the-art video generation methods, primarily focusing on text-to-video generation, tend to produce video clips with minimal motions despite maintaining high fidelity. We argue that relying solely on text instructions is insufficient and suboptimal for video generation. In this paper, we introduce PixelDance, a novel approach based on diffusion models that incorporates image instructions for both the first and last frames in conjunction with text instructions for video generation. Comprehensive experimental results demonstrate that PixelDance trained with public data exhibits significantly better proficiency in synthesizing videos with complex scenes and intricate motions, setting a new standard for video generation.

Guoqiang Wei, Cuiling Lan, Wenjun Zeng et al.

Unsupervised domain adaptive classifcation intends to improve the classifcation performance on unlabeled target domain. To alleviate the adverse effect of domain shift, many approaches align the source and target domains in the feature space. However, a feature is usually taken as a whole for alignment without explicitly making domain alignment proactively serve the classifcation task, leading to sub-optimal solution. In this paper, we propose an effective Task-oriented Alignment (ToAlign) for unsupervised domain adaptation (UDA). We study what features should be aligned across domains and propose to make the domain alignment proactively serve classifcation by performing feature decomposition and alignment under the guidance of the prior knowledge induced from the classifcation task itself. Particularly, we explicitly decompose a feature in the source domain into a task-related/discriminative feature that should be aligned, and a task-irrelevant feature that should be avoided/ignored, based on the classifcation meta-knowledge. Extensive experimental results on various benchmarks (e.g., Offce-Home, Visda-2017, and DomainNet) under different domain adaptation settings demonstrate the effectiveness of ToAlign which helps achieve the state-of-the-art performance. The code is publicly available at https://github.com/microsoft/UDA

Xin Jin, Cuiling Lan, Wenjun Zeng et al.

Person re-identification (reID) aims to match person images to retrieve the ones with the same identity. This is a challenging task, as the images to be matched are generally semantically misaligned due to the diversity of human poses and capture viewpoints, incompleteness of the visible bodies (due to occlusion), etc. In this paper, we propose a framework that drives the reID network to learn semantics-aligned feature representation through delicate supervision designs. Specifically, we build a Semantics Aligning Network (SAN) which consists of a base network as encoder (SA-Enc) for re-ID, and a decoder (SA-Dec) for reconstructing/regressing the densely semantics aligned full texture image. We jointly train the SAN under the supervisions of person re-identification and aligned texture generation. Moreover, at the decoder, besides the reconstruction loss, we add Triplet ReID constraints over the feature maps as the perceptual losses. The decoder is discarded in the inference and thus our scheme is computationally efficient. Ablation studies demonstrate the effectiveness of our design. We achieve the state-of-the-art performances on the benchmark datasets CUHK03, Market1501, MSMT17, and the partial person reID dataset Partial REID. Code for our proposed method is available at: https://github.com/microsoft/Semantics-Aligned-Representation-Learning-for-Person-Re-identification.

Jiawei Wang, Yuchen Zhang, Jiaxin Zou et al.

Generating rich and controllable motion is a pivotal challenge in video synthesis. We propose Boximator, a new approach for fine-grained motion control. Boximator introduces two constraint types: hard box and soft box. Users select objects in the conditional frame using hard boxes and then use either type of boxes to roughly or rigorously define the object's position, shape, or motion path in future frames. Boximator functions as a plug-in for existing video diffusion models. Its training process preserves the base model's knowledge by freezing the original weights and training only the control module. To address training challenges, we introduce a novel self-tracking technique that greatly simplifies the learning of box-object correlations. Empirically, Boximator achieves state-of-the-art video quality (FVD) scores, improving on two base models, and further enhanced after incorporating box constraints. Its robust motion controllability is validated by drastic increases in the bounding box alignment metric. Human evaluation also shows that users favor Boximator generation results over the base model.

Hongcan Guo, Qinyu Zhao, Yian Zhao et al.

Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.

Yu Gao, Haoyuan Guo, Tuyen Hoang et al.

Notable breakthroughs in diffusion modeling have propelled rapid improvements in video generation, yet current foundational model still face critical challenges in simultaneously balancing prompt following, motion plausibility, and visual quality. In this report, we introduce Seedance 1.0, a high-performance and inference-efficient video foundation generation model that integrates several core technical improvements: (i) multi-source data curation augmented with precision and meaningful video captioning, enabling comprehensive learning across diverse scenarios; (ii) an efficient architecture design with proposed training paradigm, which allows for natively supporting multi-shot generation and jointly learning of both text-to-video and image-to-video tasks. (iii) carefully-optimized post-training approaches leveraging fine-grained supervised fine-tuning, and video-specific RLHF with multi-dimensional reward mechanisms for comprehensive performance improvements; (iv) excellent model acceleration achieving ~10x inference speedup through multi-stage distillation strategies and system-level optimizations. Seedance 1.0 can generate a 5-second video at 1080p resolution only with 41.4 seconds (NVIDIA-L20). Compared to state-of-the-art video generation models, Seedance 1.0 stands out with high-quality and fast video generation having superior spatiotemporal fluidity with structural stability, precise instruction adherence in complex multi-subject contexts, native multi-shot narrative coherence with consistent subject representation.

Chen Hou, Guoqiang Wei, Zhibo Chen

Diffusion models have attained remarkable success in the domains of image generation and editing. It is widely recognized that employing larger inversion and denoising steps in diffusion model leads to improved image reconstruction quality. However, the editing performance of diffusion models tends to be no more satisfactory even with increasing denoising steps. The deficiency in editing could be attributed to the conditional Markovian property of the editing process, where errors accumulate throughout denoising steps. To tackle this challenge, we first propose an innovative framework where a rectifier module is incorporated to modulate diffusion model weights with residual features, thereby providing compensatory information to bridge the fidelity gap. Furthermore, we introduce a novel learning paradigm aimed at minimizing error propagation during the editing process, which trains the editing procedure in a manner similar to denoising score-matching. Extensive experiments demonstrate that our proposed framework and training strategy achieve high-fidelity reconstruction and editing results across various levels of denoising steps, meanwhile exhibits exceptional performance in terms of both quantitative metric and qualitative assessments. Moreover, we explore our model's generalization through several applications like image-to-image translation and out-of-domain image editing.

Team Seedance, Heyi Chen, Siyan Chen et al.

Recent strides in video generation have paved the way for unified audio-visual generation. In this work, we present Seedance 1.5 pro, a foundational model engineered specifically for native, joint audio-video generation. Leveraging a dual-branch Diffusion Transformer architecture, the model integrates a cross-modal joint module with a specialized multi-stage data pipeline, achieving exceptional audio-visual synchronization and superior generation quality. To ensure practical utility, we implement meticulous post-training optimizations, including Supervised Fine-Tuning (SFT) on high-quality datasets and Reinforcement Learning from Human Feedback (RLHF) with multi-dimensional reward models. Furthermore, we introduce an acceleration framework that boosts inference speed by over 10X. Seedance 1.5 pro distinguishes itself through precise multilingual and dialect lip-syncing, dynamic cinematic camera control, and enhanced narrative coherence, positioning it as a robust engine for professional-grade content creation. Seedance 1.5 pro is now accessible on Volcano Engine at https://console.volcengine.com/ark/region:ark+cn-beijing/experience/vision?type=GenVideo.

Xin Li, Zhizheng Zhang, Guoqiang Wei et al.

Confounders in deep learning are in general detrimental to model's generalization where they infiltrate feature representations. Therefore, learning causal features that are free of interference from confounders is important. Most previous causal learning based approaches employ back-door criterion to mitigate the adverse effect of certain specific confounder, which require the explicit identification of confounder. However, in real scenarios, confounders are typically diverse and difficult to be identified. In this paper, we propose a novel Confounder Identification-free Causal Visual Feature Learning (CICF) method, which obviates the need for identifying confounders. CICF models the interventions among different samples based on front-door criterion, and then approximates the global-scope intervening effect upon the instance-level interventions from the perspective of optimization. In this way, we aim to find a reliable optimization direction, which avoids the intervening effects of confounders, to learn causal features. Furthermore, we uncover the relation between CICF and the popular meta-learning strategy MAML, and provide an interpretation of why MAML works from the theoretical perspective of causal learning for the first time. Thanks to the effective learning of causal features, our CICF enables models to have superior generalization capability. Extensive experiments on domain generalization benchmark datasets demonstrate the effectiveness of our CICF, which achieves the state-of-the-art performance.

Guoqiang Wei, Cuiling Lan, Wenjun Zeng et al.

For unsupervised domain adaptation (UDA), to alleviate the effect of domain shift, many approaches align the source and target domains in the feature space by adversarial learning or by explicitly aligning their statistics. However, the optimization objective of such domain alignment is generally not coordinated with that of the object classification task itself such that their descent directions for optimization may be inconsistent. This will reduce the effectiveness of domain alignment in improving the performance of UDA. In this paper, we aim to study and alleviate the optimization inconsistency problem between the domain alignment and classification tasks. We address this by proposing an effective meta-optimization based strategy dubbed MetaAlign, where we treat the domain alignment objective and the classification objective as the meta-train and meta-test tasks in a meta-learning scheme. MetaAlign encourages both tasks to be optimized in a coordinated way, which maximizes the inner product of the gradients of the two tasks during training. Experimental results demonstrate the effectiveness of our proposed method on top of various alignment-based baseline approaches, for tasks of object classification and object detection. MetaAlign helps achieve the state-of-the-art performance.

Guoqiang Wei, Cuiling Lan, Wenjun Zeng et al.

The recent success of deep networks has significantly advanced 3D human pose estimation from 2D images. The diversity of capturing viewpoints and the flexibility of the human poses, however, remain some significant challenges. In this paper, we propose a view invariant 3D human pose estimation module to alleviate the effects of viewpoint diversity. The framework consists of a base network, which provides an initial estimation of a 3D pose, a view-invariant hierarchical correction network (VI-HC) on top of that to learn the 3D pose refinement under consistent views, and a view-invariant discriminative network (VID) to enforce high-level constraints over body configurations. In VI-HC, the initial 3D pose inputs are automatically transformed to consistent views for further refinements at the global body and local body parts level, respectively. For the VID, under consistent viewpoints, we use adversarial learning to differentiate between estimated poses and real poses to avoid implausible 3D poses. Experimental results demonstrate that the consistent viewpoints can dramatically enhance the performance. Our module shows robustness for different 3D pose base networks and achieves a significant improvement (about 9%) over a powerful baseline on the public 3D pose estimation benchmark Human3.6M.