Size Wu

Size Wu, Wenwei Zhang, Sheng Jin et al.

Pre-trained vision-language models (VLMs) learn to align vision and language representations on large-scale datasets, where each image-text pair usually contains a bag of semantic concepts. However, existing open-vocabulary object detectors only align region embeddings individually with the corresponding features extracted from the VLMs. Such a design leaves the compositional structure of semantic concepts in a scene under-exploited, although the structure may be implicitly learned by the VLMs. In this work, we propose to align the embedding of bag of regions beyond individual regions. The proposed method groups contextually interrelated regions as a bag. The embeddings of regions in a bag are treated as embeddings of words in a sentence, and they are sent to the text encoder of a VLM to obtain the bag-of-regions embedding, which is learned to be aligned to the corresponding features extracted by a frozen VLM. Applied to the commonly used Faster R-CNN, our approach surpasses the previous best results by 4.6 box AP50 and 2.8 mask AP on novel categories of open-vocabulary COCO and LVIS benchmarks, respectively. Code and models are available at https://github.com/wusize/ovdet.

Size Wu, Wenwei Zhang, Lumin Xu et al.

Open-vocabulary dense prediction tasks including object detection and image segmentation have been advanced by the success of Contrastive Language-Image Pre-training (CLIP). CLIP models, particularly those incorporating vision transformers (ViTs), have exhibited remarkable generalization ability in zero-shot image classification. However, when transferring the vision-language alignment of CLIP from global image representation to local region representation for the open-vocabulary dense prediction tasks, CLIP ViTs suffer from the domain shift from full images to local image regions. In this paper, we embark on an in-depth analysis of the region-language alignment in CLIP models, which is essential for downstream open-vocabulary dense prediction tasks. Subsequently, we propose an approach named CLIPSelf, which adapts the image-level recognition ability of CLIP ViT to local image regions without needing any region-text pairs. CLIPSelf empowers ViTs to distill itself by aligning a region representation extracted from its dense feature map with the image-level representation of the corresponding image crop. With the enhanced CLIP ViTs, we achieve new state-of-the-art performance on open-vocabulary object detection, semantic segmentation, and panoptic segmentation across various benchmarks. Models and code are released at https://github.com/wusize/CLIPSelf.

Shilin Xu, Xiangtai Li, Size Wu et al.

Open-vocabulary object detection (OVOD) aims to detect the objects beyond the set of classes observed during training. This work introduces a straightforward and efficient strategy that utilizes pre-trained vision-language models (VLM), like CLIP, to identify potential novel classes through zero-shot classification. Previous methods use a class-agnostic region proposal network to detect object proposals and consider the proposals that do not match the ground truth as background. Unlike these methods, our method will select a subset of proposals that will be considered as background during the training. Then, we treat them as novel classes during training. We refer to this approach as the self-training strategy, which enhances recall and accuracy for novel classes without requiring extra annotations, datasets, and re-training. Compared to previous pseudo methods, our approach does not require re-training and offline labeling processing, which is more efficient and effective in one-shot training. Empirical evaluations on three datasets, including LVIS, V3Det, and COCO, demonstrate significant improvements over the baseline performance without incurring additional parameters or computational costs during inference. In addition, we also apply our method to various baselines. In particular, compared with the previous method, F-VLM, our method achieves a 1.7% improvement on the LVIS dataset. Combined with the recent method CLIPSelf, our method also achieves 46.7 novel class AP on COCO without introducing extra data for pertaining. We also achieve over 6.5% improvement over the F-VLM baseline in the recent challenging V3Det dataset. We release our code and models at https://github.com/xushilin1/dst-det.

Ran Zhao, Sheng Jin, Size Wu et al.

Recent text-to-image (T2I) models have demonstrated impressive capabilities in photorealistic synthesis and instruction following. However, their reliability in knowledge-intensive settings remains largely unexplored. Unlike natural image generation, knowledge visualization requires not only semantic alignment but also strict adherence to domain knowledge, structural constraints, and symbolic conventions, exposing a critical gap between visual plausibility and scientific correctness. To systematically study this problem, we introduce KVBench, a curriculum-grounded benchmark for evaluating knowledge-intensive T2I generation. KVBench covers six senior high-school subjects: Biology, Chemistry, Geography, History, Mathematics, and Physics. The benchmark consists of 1,800 expert-curated prompts derived from over 30 authoritative textbooks. Using this benchmark, we evaluate 14 state-of-the-art open- and closed-source models, revealing substantial deficiencies in logical reasoning, symbolic precision, and multilingual robustness, with open-source models consistently underperforming proprietary systems. To address these limitations, we further propose KE-Check, a two-stage framework that improves scientific fidelity via (1) Knowledge Elaboration for structured prompt enrichment, and (2) Checklist-Guided Refinement for explicit constraint enforcement through violation identification and constraint-guided editing. KE-Check effectively mitigates scientific hallucinations, narrowing the performance gap between open-source and leading closed-source models. Data and codes are publicly available at https://github.com/zhaoran66/KVBench.

Dianyi Wang, Chaofan Ma, Feng Han et al.

Unified multimodal models often struggle with complex synthesis tasks that demand deep reasoning, and typically treat text-to-image generation and image editing as isolated capabilities rather than interconnected reasoning steps. To address this, we propose UniReason, a unified framework that harmonizes these two tasks through a dual reasoning paradigm. We formulate generation as world knowledge-enhanced planning to inject implicit constraints, and leverage editing capabilities for fine-grained visual refinement to further correct visual errors via self-reflection. This approach unifies generation and editing within a shared representation, mirroring the human cognitive process of planning followed by refinement. We support this framework by systematically constructing a large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains (e.g., cultural commonsense, physics, etc.) for planning, alongside an agent-generated corpus for visual self-correction. Extensive experiments demonstrate that UniReason achieves advanced performance on reasoning-intensive benchmarks such as WISE, KrisBench and UniREditBench, while maintaining superior general synthesis capabilities.

Hongyang Wei, Hongbo Liu, Zidong Wang et al.

The recent surge in popularity of Nano-Banana and Seedream 4.0 underscores the community's strong interest in multi-image composition tasks. Compared to single-image editing, multi-image composition presents significantly greater challenges in terms of consistency and quality, yet existing models have not disclosed specific methodological details for achieving high-quality fusion. Through statistical analysis, we identify Human-Object Interaction (HOI) as the most sought-after category by the community. We therefore systematically analyze and implement a state-of-the-art solution for multi-image composition with a primary focus on HOI-centric tasks. We present Skywork UniPic 3.0, a unified multimodal framework that integrates single-image editing and multi-image composition. Our model supports an arbitrary (1~6) number and resolution of input images, as well as arbitrary output resolutions (within a total pixel budget of 1024x1024). To address the challenges of multi-image composition, we design a comprehensive data collection, filtering, and synthesis pipeline, achieving strong performance with only 700K high-quality training samples. Furthermore, we introduce a novel training paradigm that formulates multi-image composition as a sequence-modeling problem, transforming conditional generation into unified sequence synthesis. To accelerate inference, we integrate trajectory mapping and distribution matching into the post-training stage, enabling the model to produce high-fidelity samples in just 8 steps and achieve a 12.5x speedup over standard synthesis sampling. Skywork UniPic 3.0 achieves state-of-the-art performance on single-image editing benchmark and surpasses both Nano-Banana and Seedream 4.0 on multi-image composition benchmark, thereby validating the effectiveness of our data pipeline and training paradigm. Code, models and dataset are publicly available.

Size Wu, Wenwei Zhang, Lumin Xu et al.

Unifying visual understanding and generation within a single multimodal framework remains a significant challenge, as the two inherently heterogeneous tasks require representations at different levels of granularity. Current approaches that utilize vector quantization (VQ) or variational autoencoders (VAE) for unified visual representation prioritize intrinsic imagery features over semantics, compromising understanding performance. In this work, we take inspiration from masked image modelling (MIM) that learns rich semantics via a mask-and-reconstruct pre-training and its successful extension to masked autoregressive (MAR) image generation. A preliminary study on the MAR encoder's representation reveals exceptional linear probing accuracy and precise feature response to visual concepts, which indicates MAR's potential for visual understanding tasks beyond its original generation role. Based on these insights, we present \emph{Harmon}, a unified autoregressive framework that harmonizes understanding and generation tasks with a shared MAR encoder. Through a three-stage training procedure that progressively optimizes understanding and generation capabilities, Harmon achieves state-of-the-art image generation results on the GenEval, MJHQ30K and WISE benchmarks while matching the performance of methods with dedicated semantic encoders (e.g., Janus) on image understanding benchmarks. Our code and models will be available at https://github.com/wusize/Harmon.

Size Wu, Wenwei Zhang, Lumin Xu et al.

Detecting objects accurately from a large or open vocabulary necessitates the vision-language alignment on region representations. However, learning such a region-text alignment by obtaining high-quality box annotations with text labels or descriptions is expensive and infeasible. In contrast, collecting image-text pairs is simpler but lacks precise object location information to associate regions with texts. In this paper, we propose a novel approach called Contrastive Language-Image Mosaic (CLIM), which leverages large-scale image-text pairs effectively for aligning region and text representations. CLIM combines multiple images into a mosaicked image and treats each image as a `pseudo region'. The feature of each pseudo region is extracted and trained to be similar to the corresponding text embedding while dissimilar from others by a contrastive loss, enabling the model to learn the region-text alignment without costly box annotations. As a generally applicable approach, CLIM consistently improves different open-vocabulary object detection methods that use caption supervision. Furthermore, CLIM can effectively enhance the region representation of vision-language models, thus providing stronger backbones for open-vocabulary object detectors. Our experimental results demonstrate that CLIM improves different baseline open-vocabulary object detectors by a large margin on both OV-COCO and OV-LVIS benchmarks. The code is available at https://github.com/wusize/CLIM.

Size Wu, Zhonghua Wu, Zerui Gong et al.

In this report, we present OpenUni, a simple, lightweight, and fully open-source baseline for unifying multimodal understanding and generation. Inspired by prevailing practices in unified model learning, we adopt an efficient training strategy that minimizes the training complexity and overhead by bridging the off-the-shelf multimodal large language models (LLMs) and diffusion models through a set of learnable queries and a light-weight transformer-based connector. With a minimalist choice of architecture, we demonstrate that OpenUni can: 1) generate high-quality and instruction-aligned images, and 2) achieve exceptional performance on standard benchmarks such as GenEval, DPG- Bench, and WISE, with only 1.1B and 3.1B activated parameters. To support open research and community advancement, we release all model weights, training code, and our curated training datasets (including 23M image-text pairs) at https://github.com/wusize/OpenUni.

Size Wu, Sheng Jin, Wenwei Zhang et al.

Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing drastic performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM -- grounding frozen off-the-shelf LMMs in human-AI conversations -- a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention mechanism of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation tokens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, F-LMM can be directly applied to complex tasks like reasoning segmentation, grounded conversation generation and visual chain-of-thought reasoning. Our code can be found at https://github.com/wusize/F-LMM.

Xiangtai Li, Haobo Yuan, Wei Li et al.

In this work, we address various segmentation tasks, each traditionally tackled by distinct or partially unified models. We propose OMG-Seg, One Model that is Good enough to efficiently and effectively handle all the segmentation tasks, including image semantic, instance, and panoptic segmentation, as well as their video counterparts, open vocabulary settings, prompt-driven, interactive segmentation like SAM, and video object segmentation. To our knowledge, this is the first model to handle all these tasks in one model and achieve satisfactory performance. We show that OMG-Seg, a transformer-based encoder-decoder architecture with task-specific queries and outputs, can support over ten distinct segmentation tasks and yet significantly reduce computational and parameter overhead across various tasks and datasets. We rigorously evaluate the inter-task influences and correlations during co-training. Code and models are available at https://github.com/lxtGH/OMG-Seg.

Zujin Guo, Size Wu, Zhongang Cai et al.

Existing interpolation methods use pre-trained video diffusion priors to generate intermediate frames between sparsely sampled keyframes. In the absence of 3D geometric guidance, these methods struggle to produce plausible results for complex, articulated human motions and offer limited control over the synthesized dynamics. In this paper, we introduce PoseFuse3D Keyframe Interpolator (PoseFuse3D-KI), a novel framework that integrates 3D human guidance signals into the diffusion process for Controllable Human-centric Keyframe Interpolation (CHKI). To provide rich spatial and structural cues for interpolation, our PoseFuse3D, a 3D-informed control model, features a novel SMPL-X encoder that transforms 3D geometry and shape into the 2D latent conditioning space, alongside a fusion network that integrates these 3D cues with 2D pose embeddings. For evaluation, we build CHKI-Video, a new dataset annotated with both 2D poses and 3D SMPL-X parameters. We show that PoseFuse3D-KI consistently outperforms state-of-the-art baselines on CHKI-Video, achieving a 9% improvement in PSNR and a 38% reduction in LPIPS. Comprehensive ablations demonstrate that our PoseFuse3D model improves interpolation fidelity.

Qingyu Shi, Size Wu, Jinbin Bai et al.

Visual tokenizers play a crucial role in diffusion models. The dimensionality of latent space governs both reconstruction fidelity and the semantic expressiveness of the latent feature. However, a fundamental trade-off is inherent between dimensionality and generation quality, constraining existing methods to low-dimensional latent spaces. Although recent works have leveraged vision foundation models to enrich the semantics of visual tokenizers and accelerate convergence, high-dimensional tokenizers still underperform their low-dimensional counterparts. In this work, we propose RecTok, which overcomes the limitations of high-dimensional visual tokenizers through two key innovations: flow semantic distillation and reconstruction--alignment distillation. Our key insight is to make the forward flow in flow matching semantically rich, which serves as the training space of diffusion transformers, rather than focusing on the latent space as in previous works. Specifically, our method distills the semantic information in VFMs into the forward flow trajectories in flow matching. And we further enhance the semantics by introducing a masked feature reconstruction loss. Our RecTok achieves superior image reconstruction, generation quality, and discriminative performance. It achieves state-of-the-art results on the gFID-50K under both with and without classifier-free guidance settings, while maintaining a semantically rich latent space structure. Furthermore, as the latent dimensionality increases, we observe consistent improvements. Code and model are available at https://shi-qingyu.github.io/rectok.github.io.

Huiqiang Sun, Liao Shen, Zhan Peng et al.

Cinematic storytelling is profoundly shaped by the artful manipulation of photographic elements such as depth of field and exposure. These effects are crucial in conveying mood and creating aesthetic appeal. However, controlling these effects in generative video models remains highly challenging, as most existing methods are restricted to camera motion control. In this paper, we propose CineCtrl, the first video cinematic editing framework that provides fine control over professional camera parameters (e.g., bokeh, shutter speed). We introduce a decoupled cross-attention mechanism to disentangle camera motion from photographic inputs, allowing fine-grained, independent control without compromising scene consistency. To overcome the shortage of training data, we develop a comprehensive data generation strategy that leverages simulated photographic effects with a dedicated real-world collection pipeline, enabling the construction of a large-scale dataset for robust model training. Extensive experiments demonstrate that our model generates high-fidelity videos with precisely controlled, user-specified photographic camera effects.

Kang Liao, Size Wu, Zhonghua Wu et al.

Camera-centric understanding and generation are two cornerstones of spatial intelligence, yet they are typically studied in isolation. We present Puffin, a unified camera-centric multimodal model that extends spatial awareness along the camera dimension. Puffin integrates language regression and diffusion-based generation to interpret and create scenes from arbitrary viewpoints. To bridge the modality gap between cameras and vision-language, we introduce a novel paradigm that treats camera as language, enabling thinking with camera. This guides the model to align spatially grounded visual cues with photographic terminology while reasoning across geometric context. Puffin is trained on Puffin-4M, a large-scale dataset of 4 million vision-language-camera triplets. We incorporate both global camera parameters and pixel-wise camera maps, yielding flexible and reliable spatial generation. Experiments demonstrate Puffin superior performance over specialized models for camera-centric generation and understanding. With instruction tuning, Puffin generalizes to diverse cross-view tasks such as spatial imagination, world exploration, and photography guidance. We will release the code, models, dataset pipeline, and benchmark to advance multimodal spatial intelligence research.

Size Wu, Sheng Jin, Wentao Liu et al.

This paper studies the task of estimating the 3D human poses of multiple persons from multiple calibrated camera views. Following the top-down paradigm, we decompose the task into two stages, i.e. person localization and pose estimation. Both stages are processed in coarse-to-fine manners. And we propose three task-specific graph neural networks for effective message passing. For 3D person localization, we first use Multi-view Matching Graph Module (MMG) to learn the cross-view association and recover coarse human proposals. The Center Refinement Graph Module (CRG) further refines the results via flexible point-based prediction. For 3D pose estimation, the Pose Regression Graph Module (PRG) learns both the multi-view geometry and structural relations between human joints. Our approach achieves state-of-the-art performance on CMU Panoptic and Shelf datasets with significantly lower computation complexity.