Minheng Ni

Shengming Yin, Chenfei Wu, Huan Yang et al. · microsoft-research, pku

In this paper, we propose NUWA-XL, a novel Diffusion over Diffusion architecture for eXtremely Long video generation. Most current work generates long videos segment by segment sequentially, which normally leads to the gap between training on short videos and inferring long videos, and the sequential generation is inefficient. Instead, our approach adopts a ``coarse-to-fine'' process, in which the video can be generated in parallel at the same granularity. A global diffusion model is applied to generate the keyframes across the entire time range, and then local diffusion models recursively fill in the content between nearby frames. This simple yet effective strategy allows us to directly train on long videos (3376 frames) to reduce the training-inference gap, and makes it possible to generate all segments in parallel. To evaluate our model, we build FlintstonesHD dataset, a new benchmark for long video generation. Experiments show that our model not only generates high-quality long videos with both global and local coherence, but also decreases the average inference time from 7.55min to 26s (by 94.26\%) at the same hardware setting when generating 1024 frames. The homepage link is \url{https://msra-nuwa.azurewebsites.net/}

Minheng Ni, Yabo Zhang, Kailai Feng et al.

Zero-shot referring image segmentation is a challenging task because it aims to find an instance segmentation mask based on the given referring descriptions, without training on this type of paired data. Current zero-shot methods mainly focus on using pre-trained discriminative models (e.g., CLIP). However, we have observed that generative models (e.g., Stable Diffusion) have potentially understood the relationships between various visual elements and text descriptions, which are rarely investigated in this task. In this work, we introduce a novel Referring Diffusional segmentor (Ref-Diff) for this task, which leverages the fine-grained multi-modal information from generative models. We demonstrate that without a proposal generator, a generative model alone can achieve comparable performance to existing SOTA weakly-supervised models. When we combine both generative and discriminative models, our Ref-Diff outperforms these competing methods by a significant margin. This indicates that generative models are also beneficial for this task and can complement discriminative models for better referring segmentation. Our code is publicly available at https://github.com/kodenii/Ref-Diff.

Xiaodong Wang, Chenfei Wu, Shengming Yin et al. · microsoft-research, pku

3D photography renders a static image into a video with appealing 3D visual effects. Existing approaches typically first conduct monocular depth estimation, then render the input frame to subsequent frames with various viewpoints, and finally use an inpainting model to fill those missing/occluded regions. The inpainting model plays a crucial role in rendering quality, but it is normally trained on out-of-domain data. To reduce the training and inference gap, we propose a novel self-supervised diffusion model as the inpainting module. Given a single input image, we automatically construct a training pair of the masked occluded image and the ground-truth image with random cycle-rendering. The constructed training samples are closely aligned to the testing instances, without the need of data annotation. To make full use of the masked images, we design a Masked Enhanced Block (MEB), which can be easily plugged into the UNet and enhance the semantic conditions. Towards real-world animation, we present a novel task: out-animation, which extends the space and time of input objects. Extensive experiments on real datasets show that our method achieves competitive results with existing SOTA methods.

Minheng Ni, Chenfei Wu, Xiaodong Wang et al. · pku

Avoiding synthesizing specific visual concepts is an essential challenge in responsible visual synthesis. However, the visual concept that needs to be avoided for responsible visual synthesis tends to be diverse, depending on the region, context, and usage scenarios. In this work, we formalize a new task, Open-vocabulary Responsible Visual Synthesis (ORES), where the synthesis model is able to avoid forbidden visual concepts while allowing users to input any desired content. To address this problem, we present a Two-stage Intervention (TIN) framework. By introducing 1) rewriting with learnable instruction through a large-scale language model (LLM) and 2) synthesizing with prompt intervention on a diffusion synthesis model, it can effectively synthesize images avoiding any concepts but following the user's query as much as possible. To evaluate on ORES, we provide a publicly available dataset, baseline models, and benchmark. Experimental results demonstrate the effectiveness of our method in reducing risks of image generation. Our work highlights the potential of LLMs in responsible visual synthesis. Our code and dataset is public available.

Minheng Ni, Zitong Huang, Kailai Feng et al.

Without the demand of training in reality, humans can easily detect a known concept simply based on its language description. Empowering deep learning with this ability undoubtedly enables the neural network to handle complex vision tasks, e.g., object detection, without collecting and annotating real images. To this end, this paper introduces a novel challenging learning paradigm Imaginary-Supervised Object Detection (ISOD), where neither real images nor manual annotations are allowed for training object detectors. To resolve this challenge, we propose ImaginaryNet, a framework to synthesize images by combining pretrained language model and text-to-image synthesis model. Given a class label, the language model is used to generate a full description of a scene with a target object, and the text-to-image model deployed to generate a photo-realistic image. With the synthesized images and class labels, weakly supervised object detection can then be leveraged to accomplish ISOD. By gradually introducing real images and manual annotations, ImaginaryNet can collaborate with other supervision settings to further boost detection performance. Experiments show that ImaginaryNet can (i) obtain about 70% performance in ISOD compared with the weakly supervised counterpart of the same backbone trained on real data, (ii) significantly improve the baseline while achieving state-of-the-art or comparable performance by incorporating ImaginaryNet with other supervision settings.

Guanghao Zhu, Zeyu Liu, Zhitian Hou et al.

Large-scale biomedical image-text datasets extracted from scientific literature provide valuable resources for medical multimodal model training. These datasets are commonly organized as image-caption pairs; however, figure captions are often short, context-dependent, and only partially informative without the surrounding article text. At the same time, large-scale automatic extraction introduces structural noise such as missing captions, residual markup, duplicated context, and incoherent multi-paragraph figure descriptions. We revisit data construction for medical multimodal continued pretraining (CPT) and present PMC-InterCPT, a context-grounded biomedical interleaved corpus that incorporates figure-referencing body text in addition to captions. Our pipeline recovers missing captions, cleans caption and context text, reconstructs coherent interleaved image-text samples, and applies LLM-supervised medical relevance and quality classifiers to filter noisy records. We further reveal strong modality imbalance in the resulting corpus and introduce a four-bucket evidence taxonomy for modality-aware resampling. Through CPT followed by supervised fine-tuning (SFT) on Qwen3.5-4B-Base, PMC-InterCPT effectively improves medical and general multimodal performance while using fewer CPT tokens than the raw source pool. The experimental results also illustrate the complementarity between the data quality and modality for medical multimodal CPT.

Minheng Ni, Chenfei Wu, Huaying Yuan et al.

With the advancement of generative models, the synthesis of different sensory elements such as music, visuals, and speech has achieved significant realism. However, the approach to generate multi-sensory outputs has not been fully explored, limiting the application on high-value scenarios such as of directing a film. Developing a movie director agent faces two major challenges: (1) Lack of parallelism and online scheduling with production steps: In the production of multi-sensory films, there are complex dependencies between different sensory elements, and the production time for each element varies. (2) Diverse needs and clear communication demands with users: Users often cannot clearly express their needs until they see a draft, which requires human-computer interaction and iteration to continually adjust and optimize the film content based on user feedback. To address these issues, we introduce AutoDirector, an interactive multi-sensory composition framework that supports long shots, special effects, music scoring, dubbing, and lip-syncing. This framework improves the efficiency of multi-sensory film production through automatic scheduling and supports the modification and improvement of interactive tasks to meet user needs. AutoDirector not only expands the application scope of human-machine collaboration but also demonstrates the potential of AI in collaborating with humans in the role of a film director to complete multi-sensory films.

Minheng Ni, Yeli Shen, Lei Zhang et al.

With recent advancements in visual synthesis, there is a growing risk of encountering images with detrimental effects, such as hate, discrimination, or privacy violations. The research on transforming harmful images into responsible ones remains unexplored. In this paper, we formulate a new task, responsible visual editing, which entails modifying specific concepts within an image to render it more responsible while minimizing changes. However, the concept that needs to be edited is often abstract, making it challenging to locate what needs to be modified and plan how to modify it. To tackle these challenges, we propose a Cognitive Editor (CoEditor) that harnesses the large multimodal model through a two-stage cognitive process: (1) a perceptual cognitive process to focus on what needs to be modified and (2) a behavioral cognitive process to strategize how to modify. To mitigate the negative implications of harmful images on research, we create a transparent and public dataset, AltBear, which expresses harmful information using teddy bears instead of humans. Experiments demonstrate that CoEditor can effectively comprehend abstract concepts within complex scenes and significantly surpass the performance of baseline models for responsible visual editing. We find that the AltBear dataset corresponds well to the harmful content found in real images, offering a consistent experimental evaluation, thereby providing a safer benchmark for future research. Moreover, CoEditor also shows great results in general editing. We release our code and dataset at https://github.com/kodenii/Responsible-Visual-Editing.

Minheng Ni, Ennan Wu, Zidong Gong et al. · microsoft-research

The widespread application of artificial intelligence (AI) in various tasks, along with frequent reports of conflicts or violations involving AI, has sparked societal concerns about interactions with AI systems. Based on Wrightsman's Philosophies of Human Nature Scale (PHNS), a scale empirically validated over decades to effectively assess individuals' attitudes toward human nature, we design the standardized psychological scale specifically targeting large language models (LLM), named the Machine-based Philosophies of Human Nature Scale (M-PHNS). By evaluating LLMs' attitudes toward human nature across six dimensions, we reveal that current LLMs exhibit a systemic lack of trust in humans, and there is a significant negative correlation between the model's intelligence level and its trust in humans. Furthermore, we propose a mental loop learning framework, which enables LLM to continuously optimize its value system during virtual interactions by constructing moral scenarios, thereby improving its attitude toward human nature. Experiments demonstrate that mental loop learning significantly enhances their trust in humans compared to persona or instruction prompts. This finding highlights the potential of human-based psychological assessments for LLM, which can not only diagnose cognitive biases but also provide a potential solution for ethical learning in artificial intelligence. We release the M-PHNS evaluation code and data at https://github.com/kodenii/M-PHNS.

Minheng Ni, Lei Zhang, Zihan Chen et al.

Unthinking execution of human instructions in robotic manipulation can lead to severe safety risks, such as poisonings, fires, and even explosions. In this paper, we present responsible robotic manipulation, which requires robots to consider potential hazards in the real-world environment while completing instructions and performing complex operations safely and efficiently. However, such scenarios in real world are variable and risky for training. To address this challenge, we propose Safety-as-policy, which includes (i) a world model to automatically generate scenarios containing safety risks and conduct virtual interactions, and (ii) a mental model to infer consequences with reflections and gradually develop the cognition of safety, allowing robots to accomplish tasks while avoiding dangers. Additionally, we create the SafeBox synthetic dataset, which includes one hundred responsible robotic manipulation tasks with different safety risk scenarios and instructions, effectively reducing the risks associated with real-world experiments. Experiments demonstrate that Safety-as-policy can avoid risks and efficiently complete tasks in both synthetic dataset and real-world experiments, significantly outperforming baseline methods. Our SafeBox dataset shows consistent evaluation results with real-world scenarios, serving as a safe and effective benchmark for future research.

Bowen Dong, Minheng Ni, Zitong Huang et al.

Multimodal hallucination in multimodal large language models (MLLMs) restricts the correctness of MLLMs. However, multimodal hallucinations are multi-sourced and arise from diverse causes. Existing benchmarks fail to adequately distinguish between perception-induced hallucinations and reasoning-induced hallucinations. This failure constitutes a significant issue and hinders the diagnosis of multimodal reasoning failures within MLLMs. To address this, we propose the {\dataset} benchmark, which isolates reasoning hallucinations by constructing questions where input images are correctly perceived by MLLMs yet reasoning errors persist. {\dataset} introduces multi-granular evaluation metrics: accuracy, factuality, and LLMs hallucination score for hallucination quantification. Our analysis reveals that (1) the model scale, data scale, and training stages significantly affect the degree of logical, fabrication, and factual hallucinations; (2) current MLLMs show no effective improvement on spatial hallucinations caused by misinterpreted spatial relationships, indicating their limited visual reasoning capabilities; and (3) question types correlate with distinct hallucination patterns, highlighting targeted challenges and potential mitigation strategies. To address these challenges, we propose {\method}, a method that combines curriculum reinforcement fine-tuning to encourage models to generate logic-consistent reasoning chains by stepwise reducing learning difficulty, and collaborative hint inference to reduce reasoning complexity. {\method} establishes a baseline on {\dataset}, and reduces the logical hallucinations in original base models.

Minheng Ni, Zhengyuan Yang, Yaowen Zhang et al.

We study professional image generation, where a model must synthesize information-dense, scientifically precise illustrations from technical descriptions rather than merely produce visually plausible pictures. To quantify the progress, we introduce ProImage-Bench, a rubric-based benchmark that targets biology schematics, engineering/patent drawings, and general scientific diagrams. For 654 figures collected from real textbooks and technical reports, we construct detailed image instructions and a hierarchy of rubrics that decompose correctness into 6,076 criteria and 44,131 binary checks. Rubrics are derived from surrounding text and reference figures using large multimodal models, and are evaluated by an automated LMM-based judge with a principled penalty scheme that aggregates sub-question outcomes into interpretable criterion scores. We benchmark several representative text-to-image models on ProImage-Bench and find that, despite strong open-domain performance, the best base model reaches only 0.791 rubric accuracy and 0.553 criterion score overall, revealing substantial gaps in fine-grained scientific fidelity. Finally, we show that the same rubrics provide actionable supervision: feeding failed checks back into an editing model for iterative refinement boosts a strong generator from 0.653 to 0.865 in rubric accuracy and from 0.388 to 0.697 in criterion score. ProImage-Bench thus offers both a rigorous diagnostic for professional image generation and a scalable signal for improving specification-faithful scientific illustrations.

Minheng Ni, Chenfei Wu, Haoyang Huang et al.

Language guided image inpainting aims to fill in the defective regions of an image under the guidance of text while keeping non-defective regions unchanged. However, the encoding process of existing models suffers from either receptive spreading of defective regions or information loss of non-defective regions, giving rise to visually unappealing inpainting results. To address the above issues, this paper proposes NÜWA-LIP by incorporating defect-free VQGAN (DF-VQGAN) with multi-perspective sequence to sequence (MP-S2S). In particular, DF-VQGAN introduces relative estimation to control receptive spreading and adopts symmetrical connections to protect information. MP-S2S further enhances visual information from complementary perspectives, including both low-level pixels and high-level tokens. Experiments show that DF-VQGAN performs more robustness than VQGAN. To evaluate the inpainting performance of our model, we built up 3 open-domain benchmarks, where NÜWA-LIP is also superior to recent strong baselines.

Libo Qin, Zhouyang Li, Wanxiang Che et al.

In a dialog system, dialog act recognition and sentiment classification are two correlative tasks to capture speakers intentions, where dialog act and sentiment can indicate the explicit and the implicit intentions separately. The dialog context information (contextual information) and the mutual interaction information are two key factors that contribute to the two related tasks. Unfortunately, none of the existing approaches consider the two important sources of information simultaneously. In this paper, we propose a Co-Interactive Graph Attention Network (Co-GAT) to jointly perform the two tasks. The core module is a proposed co-interactive graph interaction layer where a cross-utterances connection and a cross-tasks connection are constructed and iteratively updated with each other, achieving to consider the two types of information simultaneously. Experimental results on two public datasets show that our model successfully captures the two sources of information and achieve the state-of-the-art performance. In addition, we find that the contributions from the contextual and mutual interaction information do not fully overlap with contextualized word representations (BERT, Roberta, XLNet).

Libo Qin, Wanxiang Che, Yangming Li et al.

In dialog system, dialog act recognition and sentiment classification are two correlative tasks to capture speakers intentions, where dialog act and sentiment can indicate the explicit and the implicit intentions separately. Most of the existing systems either treat them as separate tasks or just jointly model the two tasks by sharing parameters in an implicit way without explicitly modeling mutual interaction and relation. To address this problem, we propose a Deep Co-Interactive Relation Network (DCR-Net) to explicitly consider the cross-impact and model the interaction between the two tasks by introducing a co-interactive relation layer. In addition, the proposed relation layer can be stacked to gradually capture mutual knowledge with multiple steps of interaction. Especially, we thoroughly study different relation layers and their effects. Experimental results on two public datasets (Mastodon and Dailydialog) show that our model outperforms the state-of-the-art joint model by 4.3% and 3.4% in terms of F1 score on dialog act recognition task, 5.7% and 12.4% on sentiment classification respectively. Comprehensive analysis empirically verifies the effectiveness of explicitly modeling the relation between the two tasks and the multi-steps interaction mechanism. Finally, we employ the Bidirectional Encoder Representation from Transformer (BERT) in our framework, which can further boost our performance in both tasks.

Libo Qin, Minheng Ni, Yue Zhang et al.

Multi-lingual contextualized embeddings, such as multilingual-BERT (mBERT), have shown success in a variety of zero-shot cross-lingual tasks. However, these models are limited by having inconsistent contextualized representations of subwords across different languages. Existing work addresses this issue by bilingual projection and fine-tuning technique. We propose a data augmentation framework to generate multi-lingual code-switching data to fine-tune mBERT, which encourages model to align representations from source and multiple target languages once by mixing their context information. Compared with the existing work, our method does not rely on bilingual sentences for training, and requires only one training process for multiple target languages. Experimental results on five tasks with 19 languages show that our method leads to significantly improved performances for all the tasks compared with mBERT.

Minheng Ni, Haoyang Huang, Lin Su et al.

We present M3P, a Multitask Multilingual Multimodal Pre-trained model that combines multilingual pre-training and multimodal pre-training into a unified framework via multitask pre-training. Our goal is to learn universal representations that can map objects occurred in different modalities or texts expressed in different languages into a common semantic space. In addition, to explicitly encourage fine-grained alignment between images and non-English languages, we also propose Multimodal Code-switched Training (MCT) to combine monolingual pre-training and multimodal pre-training via a code-switch strategy. Experiments are performed on the multilingual image retrieval task across two benchmark datasets, including MSCOCO and Multi30K. M3P can achieve comparable results for English and new state-of-the-art results for non-English languages.

Libo Qin, Minheng Ni, Yue Zhang et al.

Spoken language understanding has been addressed as a supervised learning problem, where a set of training data is available for each domain. However, annotating data for each domain is both financially costly and non-scalable so we should fully utilize information across all domains. One existing approach solves the problem by conducting multi-domain learning, using shared parameters for joint training across domains. We propose to improve the parameterization of this method by using domain-specific and task-specific model parameters to improve knowledge learning and transfer. Experiments on 5 domains show that our model is more effective for multi-domain SLU and obtain the best results. In addition, we show its transferability by outperforming the prior best model by 12.4\% when adapting to a new domain with little data.