Tsu-Jui Fu

Tsu-Jui Fu, Wenze Hu, Xianzhi Du et al. · microsoft-research

Instruction-based image editing improves the controllability and flexibility of image manipulation via natural commands without elaborate descriptions or regional masks. However, human instructions are sometimes too brief for current methods to capture and follow. Multimodal large language models (MLLMs) show promising capabilities in cross-modal understanding and visual-aware response generation via LMs. We investigate how MLLMs facilitate edit instructions and present MLLM-Guided Image Editing (MGIE). MGIE learns to derive expressive instructions and provides explicit guidance. The editing model jointly captures this visual imagination and performs manipulation through end-to-end training. We evaluate various aspects of Photoshop-style modification, global photo optimization, and local editing. Extensive experimental results demonstrate that expressive instructions are crucial to instruction-based image editing, and our MGIE can lead to a notable improvement in automatic metrics and human evaluation while maintaining competitive inference efficiency.

Tsu-Jui Fu, Linjie Li, Zhe Gan et al. · microsoft-research, uw

Masked visual modeling (MVM) has been recently proven effective for visual pre-training. While similar reconstructive objectives on video inputs (e.g., masked frame modeling) have been explored in video-language (VidL) pre-training, previous studies fail to find a truly effective MVM strategy that can largely benefit the downstream performance. In this work, we systematically examine the potential of MVM in the context of VidL learning. Specifically, we base our study on a fully end-to-end VIdeO-LanguagE Transformer (VIOLET), where the supervision from MVM training can be backpropagated to the video pixel space. In total, eight different reconstructive targets of MVM are explored, from low-level pixel values and oriented gradients to high-level depth maps, optical flow, discrete visual tokens, and latent visual features. We conduct comprehensive experiments and provide insights into the factors leading to effective MVM training, resulting in an enhanced model VIOLETv2. Empirically, we show VIOLETv2 pre-trained with MVM objective achieves notable improvements on 13 VidL benchmarks, ranging from video question answering, video captioning, to text-to-video retrieval.

Weixi Feng, Xuehai He, Tsu-Jui Fu et al. · ibm-research

Large-scale diffusion models have achieved state-of-the-art results on text-to-image synthesis (T2I) tasks. Despite their ability to generate high-quality yet creative images, we observe that attribution-binding and compositional capabilities are still considered major challenging issues, especially when involving multiple objects. In this work, we improve the compositional skills of T2I models, specifically more accurate attribute binding and better image compositions. To do this, we incorporate linguistic structures with the diffusion guidance process based on the controllable properties of manipulating cross-attention layers in diffusion-based T2I models. We observe that keys and values in cross-attention layers have strong semantic meanings associated with object layouts and content. Therefore, we can better preserve the compositional semantics in the generated image by manipulating the cross-attention representations based on linguistic insights. Built upon Stable Diffusion, a SOTA T2I model, our structured cross-attention design is efficient that requires no additional training samples. We achieve better compositional skills in qualitative and quantitative results, leading to a 5-8% advantage in head-to-head user comparison studies. Lastly, we conduct an in-depth analysis to reveal potential causes of incorrect image compositions and justify the properties of cross-attention layers in the generation process.

Xuehai He, Diji Yang, Weixi Feng et al. · ibm-research

Prompt tuning is a new few-shot transfer learning technique that only tunes the learnable prompt for pre-trained vision and language models such as CLIP. However, existing prompt tuning methods tend to learn spurious or entangled representations, which leads to poor generalization to unseen concepts. Towards non-spurious and efficient prompt learning from limited examples, this paper presents a novel \underline{\textbf{C}}ounterfactual \underline{\textbf{P}}rompt \underline{\textbf{L}}earning (CPL) method for vision and language models, which simultaneously employs counterfactual generation and contrastive learning in a joint optimization framework. Particularly, CPL constructs counterfactual by identifying minimal non-spurious feature change between semantically-similar positive and negative samples that causes concept change, and learns more generalizable prompt representation from both factual and counterfactual examples via contrastive learning. Extensive experiments demonstrate that CPL can obtain superior few-shot performance on different vision and language tasks than previous prompt tuning methods on CLIP. On image classification, we achieve 3.55\% average relative improvement on unseen classes across seven datasets; on image-text retrieval and visual question answering, we gain up to 4.09\% and 25.08\% relative improvements across three few-shot scenarios on unseen test sets respectively.

Raphael Schumann, Wanrong Zhu, Weixi Feng et al.

Incremental decision making in real-world environments is one of the most challenging tasks in embodied artificial intelligence. One particularly demanding scenario is Vision and Language Navigation~(VLN) which requires visual and natural language understanding as well as spatial and temporal reasoning capabilities. The embodied agent needs to ground its understanding of navigation instructions in observations of a real-world environment like Street View. Despite the impressive results of LLMs in other research areas, it is an ongoing problem of how to best connect them with an interactive visual environment. In this work, we propose VELMA, an embodied LLM agent that uses a verbalization of the trajectory and of visual environment observations as contextual prompt for the next action. Visual information is verbalized by a pipeline that extracts landmarks from the human written navigation instructions and uses CLIP to determine their visibility in the current panorama view. We show that VELMA is able to successfully follow navigation instructions in Street View with only two in-context examples. We further finetune the LLM agent on a few thousand examples and achieve 25%-30% relative improvement in task completion over the previous state-of-the-art for two datasets.

Tsu-Jui Fu, Licheng Yu, Ning Zhang et al.

Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC.

Weixi Feng, Tsu-Jui Fu, Yujie Lu et al.

Vision-and-Language Navigation (VLN) is a task to guide an embodied agent moving to a target position using language instructions. Despite the significant performance improvement, the wide use of fine-grained instructions fails to characterize more practical linguistic variations in reality. To fill in this gap, we introduce a new setting, namely Underspecified vision-and-Language Navigation (ULN), and associated evaluation datasets. ULN evaluates agents using multi-level underspecified instructions instead of purely fine-grained or coarse-grained, which is a more realistic and general setting. As a primary step toward ULN, we propose a VLN framework that consists of a classification module, a navigation agent, and an Exploitation-to-Exploration (E2E) module. Specifically, we propose to learn Granularity Specific Sub-networks (GSS) for the agent to ground multi-level instructions with minimal additional parameters. Then, our E2E module estimates grounding uncertainty and conducts multi-step lookahead exploration to improve the success rate further. Experimental results show that existing VLN models are still brittle to multi-level language underspecification. Our framework is more robust and outperforms the baselines on ULN by ~10% relative success rate across all levels.

Yujie Lu, Xiujun Li, Tsu-Jui Fu et al.

The rapid progress in Multimodal Large Language Models (MLLMs) has significantly advanced their ability to process and understand complex visual and textual information. However, the integration of multiple images and extensive textual contexts remains a challenge due to the inherent limitation of the models' capacity to handle long input sequences efficiently. In this paper, we introduce SEEKER, a multimodal large language model designed to tackle this issue. SEEKER aims to optimize the compact encoding of long text by compressing the text sequence into the visual pixel space via images, enabling the model to handle long text within a fixed token-length budget efficiently. Our empirical experiments on six long-context multimodal tasks demonstrate that SEEKER can leverage fewer image tokens to convey the same amount of textual information compared with the OCR-based approach, and is more efficient in understanding long-form multimodal input and generating long-form textual output, outperforming all existing proprietary and open-source MLLMs by large margins.

Peng-Hsuan Li, Tsu-Jui Fu, Wei-Yun Ma

BiLSTM has been prevalently used as a core module for NER in a sequence-labeling setup. State-of-the-art approaches use BiLSTM with additional resources such as gazetteers, language-modeling, or multi-task supervision to further improve NER. This paper instead takes a step back and focuses on analyzing problems of BiLSTM itself and how exactly self-attention can bring improvements. We formally show the limitation of (CRF-)BiLSTM in modeling cross-context patterns for each word -- the XOR limitation. Then, we show that two types of simple cross-structures -- self-attention and Cross-BiLSTM -- can effectively remedy the problem. We test the practical impacts of the deficiency on real-world NER datasets, OntoNotes 5.0 and WNUT 2017, with clear and consistent improvements over the baseline, up to 8.7% on some of the multi-token entity mentions. We give in-depth analyses of the improvements across several aspects of NER, especially the identification of multi-token mentions. This study should lay a sound foundation for future improvements on sequence-labeling NER. (Source codes: https://github.com/jacobvsdanniel/cross-ner)

Haotian Zhang, Haoxuan You, Philipp Dufter et al.

While Ferret seamlessly integrates regional understanding into the Large Language Model (LLM) to facilitate its referring and grounding capability, it poses certain limitations: constrained by the pre-trained fixed visual encoder and failed to perform well on broader tasks. In this work, we unveil Ferret-v2, a significant upgrade to Ferret, with three key designs. (1) Any resolution grounding and referring: A flexible approach that effortlessly handles higher image resolution, improving the model's ability to process and understand images in greater detail. (2) Multi-granularity visual encoding: By integrating the additional DINOv2 encoder, the model learns better and diverse underlying contexts for global and fine-grained visual information. (3) A three-stage training paradigm: Besides image-caption alignment, an additional stage is proposed for high-resolution dense alignment before the final instruction tuning. Experiments show that Ferret-v2 provides substantial improvements over Ferret and other state-of-the-art methods, thanks to its high-resolution scaling and fine-grained visual processing.

Xiaoyu Wu, Yifei Wang, Tsu-Jui Fu et al.

We study outlier tokens in Diffusion Transformers (DiTs) for image generation. Prior work has shown that Vision Transformers (ViTs) can produce a small number of high-norm tokens that attract disproportionate attention while carrying limited local information, but their role in generative models remains underexplored. We show that this phenomenon appears in both the encoder and denoiser of modern Representation Autoencoder (RAE)-DiT pipelines: pretrained ViT encoders can produce outlier representations, and DiTs themselves can develop internal outlier tokens, especially in intermediate layers. Moreover, simply masking high-norm tokens does not improve performance, indicating that the problem is not only caused by a few extreme values, but is more closely related to corrupted local patch semantics. To address this issue, we introduce Dual-Stage Registers (DSR), a register-based intervention for both components: trained registers when available, recursive test-time registers otherwise, and diffusion registers for the denoiser. Across ImageNet and large-scale text-to-image generation, these interventions consistently reduce outlier artifacts and improve generation quality. Our results highlight outlier-token control as an important ingredient in building stronger DiTs.

Zongyu Lin, Wei Liu, Chen Chen et al.

The field of video generation has made remarkable advancements, yet there remains a pressing need for a clear, systematic recipe that can guide the development of robust and scalable models. In this work, we present a comprehensive study that systematically explores the interplay of model architectures, training recipes, and data curation strategies, culminating in a simple and scalable text-image-conditioned video generation method, named STIV. Our framework integrates image condition into a Diffusion Transformer (DiT) through frame replacement, while incorporating text conditioning via a joint image-text conditional classifier-free guidance. This design enables STIV to perform both text-to-video (T2V) and text-image-to-video (TI2V) tasks simultaneously. Additionally, STIV can be easily extended to various applications, such as video prediction, frame interpolation, multi-view generation, and long video generation, etc. With comprehensive ablation studies on T2I, T2V, and TI2V, STIV demonstrate strong performance, despite its simple design. An 8.7B model with 512 resolution achieves 83.1 on VBench T2V, surpassing both leading open and closed-source models like CogVideoX-5B, Pika, Kling, and Gen-3. The same-sized model also achieves a state-of-the-art result of 90.1 on VBench I2V task at 512 resolution. By providing a transparent and extensible recipe for building cutting-edge video generation models, we aim to empower future research and accelerate progress toward more versatile and reliable video generation solutions.

Tsu-Jui Fu, Yusu Qian, Chen Chen et al.

Text-to-Image (T2I) diffusion models have shown impressive results in generating visually compelling images following user prompts. Building on this, various methods further fine-tune the pre-trained T2I model for specific tasks. However, this requires separate model architectures, training designs, and multiple parameter sets to handle different tasks. In this paper, we introduce UniVG, a generalist diffusion model capable of supporting a diverse range of image generation tasks with a single set of weights. UniVG treats multi-modal inputs as unified conditions to enable various downstream applications, ranging from T2I generation, inpainting, instruction-based editing, identity-preserving generation, and layout-guided generation, to depth estimation and referring segmentation. Through comprehensive empirical studies on data mixing and multi-task training, we provide detailed insights into the training processes and decisions that inform our final designs. For example, we show that T2I generation and other tasks, such as instruction-based editing, can coexist without performance trade-offs, while auxiliary tasks like depth estimation and referring segmentation enhance image editing. Notably, our model can even outperform some task-specific models on their respective benchmarks, marking a significant step towards a unified image generation model.

Chen Chen, Rui Qian, Wenze Hu et al.

In this work, we empirically study Diffusion Transformers (DiTs) for text-to-image generation, focusing on architectural choices, text-conditioning strategies, and training protocols. We evaluate a range of DiT-based architectures--including PixArt-style and MMDiT variants--and compare them with a standard DiT variant which directly processes concatenated text and noise inputs. Surprisingly, our findings reveal that the performance of standard DiT is comparable with those specialized models, while demonstrating superior parameter-efficiency, especially when scaled up. Leveraging the layer-wise parameter sharing strategy, we achieve a further reduction of 66% in model size compared to an MMDiT architecture, with minimal performance impact. Building on an in-depth analysis of critical components such as text encoders and Variational Auto-Encoders (VAEs), we introduce DiT-Air and DiT-Air-Lite. With supervised and reward fine-tuning, DiT-Air achieves state-of-the-art performance on GenEval and T2I CompBench, while DiT-Air-Lite remains highly competitive, surpassing most existing models despite its compact size.

Long Lian, Sida Wang, Felix Juefei-Xu et al.

Scaling inference-time computation has enabled Large Language Models (LLMs) to achieve strong reasoning performance, but inherently sequential decoding leads to substantial latency, especially on complex tasks. Recent work on adaptive parallel reasoning aims to improve inference efficiency by decomposing the problem-solving process into concurrent reasoning threads when beneficial. However, existing methods on realistic tasks are either limited to supervised behavior cloning or exhibit significant accuracy drops compared to widely-used sequential long chain-of-thought (CoT) baselines. Moreover, many require customized inference engines, complicating deployment. We introduce ThreadWeaver, a framework for adaptive parallel reasoning that achieves accuracy on par with popular sequential reasoning models of comparable size while significantly reducing inference latency. ThreadWeaver's performance stems from three key innovations: 1) a two-stage parallel trajectory generator that produces large-scale, high-quality CoT data with parallel annotations for supervised fine-tuning; 2) a trie-based training-inference co-design that enables parallel reasoning on any off-the-shelf autoregressive inference engine without modifying position embeddings or KV caches; and 3) a parallelization-aware reinforcement learning framework that teaches the model to balance accuracy with effective parallelization. Across six challenging mathematical reasoning benchmarks, ThreadWeaver trained atop Qwen3-8B achieves accuracy comparable to cutting-edge sequential reasoning models (71.9% on average and 79.9% on AIME24) while delivering up to 1.53x average speedup in token latency, establishing a new Pareto frontier between accuracy and efficiency.

Chen Chen, Pengsheng Guo, Liangchen Song et al.

Conditional generative modeling aims to learn a conditional data distribution from samples containing data-condition pairs. For this, diffusion and flow-based methods have attained compelling results. These methods use a learned (flow) model to transport an initial standard Gaussian noise that ignores the condition to the conditional data distribution. The model is hence required to learn both mass transport and conditional injection. To ease the demand on the model, we propose Condition-Aware Reparameterization for Flow Matching (CAR-Flow) -- a lightweight, learned shift that conditions the source, the target, or both distributions. By relocating these distributions, CAR-Flow shortens the probability path the model must learn, leading to faster training in practice. On low-dimensional synthetic data, we visualize and quantify the effects of CAR-Flow. On higher-dimensional natural image data (ImageNet-256), equipping SiT-XL/2 with CAR-Flow reduces FID from 2.07 to 1.68, while introducing less than 0.6% additional parameters.

Weixi Feng, Jiachen Li, Michael Saxon et al.

Video generation has many unique challenges beyond those of image generation. The temporal dimension introduces extensive possible variations across frames, over which consistency and continuity may be violated. In this study, we move beyond evaluating simple actions and argue that generated videos should incorporate the emergence of new concepts and their relation transitions like in real-world videos as time progresses. To assess the Temporal Compositionality of video generation models, we propose TC-Bench, a benchmark of meticulously crafted text prompts, corresponding ground truth videos, and robust evaluation metrics. The prompts articulate the initial and final states of scenes, effectively reducing ambiguities for frame development and simplifying the assessment of transition completion. In addition, by collecting aligned real-world videos corresponding to the prompts, we expand TC-Bench's applicability from text-conditional models to image-conditional ones that can perform generative frame interpolation. We also develop new metrics to measure the completeness of component transitions in generated videos, which demonstrate significantly higher correlations with human judgments than existing metrics. Our comprehensive experimental results reveal that most video generators achieve less than 20% of the compositional changes, highlighting enormous space for future improvement. Our analysis indicates that current video generation models struggle to interpret descriptions of compositional changes and synthesize various components across different time steps.

Jing Gu, Yilin Wang, Nanxuan Zhao et al.

In an era where images and visual content dominate our digital landscape, the ability to manipulate and personalize these images has become a necessity. Envision seamlessly substituting a tabby cat lounging on a sunlit window sill in a photograph with your own playful puppy, all while preserving the original charm and composition of the image. We present Photoswap, a novel approach that enables this immersive image editing experience through personalized subject swapping in existing images. Photoswap first learns the visual concept of the subject from reference images and then swaps it into the target image using pre-trained diffusion models in a training-free manner. We establish that a well-conceptualized visual subject can be seamlessly transferred to any image with appropriate self-attention and cross-attention manipulation, maintaining the pose of the swapped subject and the overall coherence of the image. Comprehensive experiments underscore the efficacy and controllability of Photoswap in personalized subject swapping. Furthermore, Photoswap significantly outperforms baseline methods in human ratings across subject swapping, background preservation, and overall quality, revealing its vast application potential, from entertainment to professional editing.

Weixi Feng, Wanrong Zhu, Tsu-jui Fu et al.

Attaining a high degree of user controllability in visual generation often requires intricate, fine-grained inputs like layouts. However, such inputs impose a substantial burden on users when compared to simple text inputs. To address the issue, we study how Large Language Models (LLMs) can serve as visual planners by generating layouts from text conditions, and thus collaborate with visual generative models. We propose LayoutGPT, a method to compose in-context visual demonstrations in style sheet language to enhance the visual planning skills of LLMs. LayoutGPT can generate plausible layouts in multiple domains, ranging from 2D images to 3D indoor scenes. LayoutGPT also shows superior performance in converting challenging language concepts like numerical and spatial relations to layout arrangements for faithful text-to-image generation. When combined with a downstream image generation model, LayoutGPT outperforms text-to-image models/systems by 20-40% and achieves comparable performance as human users in designing visual layouts for numerical and spatial correctness. Lastly, LayoutGPT achieves comparable performance to supervised methods in 3D indoor scene synthesis, demonstrating its effectiveness and potential in multiple visual domains.

Tsu-Jui Fu, Wenhan Xiong, Yixin Nie et al.

3D human modeling has been widely used for engaging interaction in gaming, film, and animation. The customization of these characters is crucial for creativity and scalability, which highlights the importance of controllability. In this work, we introduce Text-guided 3D Human Generation (\texttt{T3H}), where a model is to generate a 3D human, guided by the fashion description. There are two goals: 1) the 3D human should render articulately, and 2) its outfit is controlled by the given text. To address this \texttt{T3H} task, we propose Compositional Cross-modal Human (CCH). CCH adopts cross-modal attention to fuse compositional human rendering with the extracted fashion semantics. Each human body part perceives relevant textual guidance as its visual patterns. We incorporate the human prior and semantic discrimination to enhance 3D geometry transformation and fine-grained consistency, enabling it to learn from 2D collections for data efficiency. We conduct evaluations on DeepFashion and SHHQ with diverse fashion attributes covering the shape, fabric, and color of upper and lower clothing. Extensive experiments demonstrate that CCH achieves superior results for \texttt{T3H} with high efficiency.

Siqi Liu, Weixi Feng, Tsu-jui Fu et al.

Making image retrieval methods practical for real-world search applications requires significant progress in dataset scales, entity comprehension, and multimodal information fusion. In this work, we introduce \textbf{E}ntity-\textbf{D}riven \textbf{I}mage \textbf{S}earch (EDIS), a challenging dataset for cross-modal image search in the news domain. EDIS consists of 1 million web images from actual search engine results and curated datasets, with each image paired with a textual description. Unlike datasets that assume a small set of single-modality candidates, EDIS reflects real-world web image search scenarios by including a million multimodal image-text pairs as candidates. EDIS encourages the development of retrieval models that simultaneously address cross-modal information fusion and matching. To achieve accurate ranking results, a model must: 1) understand named entities and events from text queries, 2) ground entities onto images or text descriptions, and 3) effectively fuse textual and visual representations. Our experimental results show that EDIS challenges state-of-the-art methods with dense entities and a large-scale candidate set. The ablation study also proves that fusing textual features with visual features is critical in improving retrieval results.

Wanrong Zhu, Xinyi Wang, Yujie Lu et al.

The field of text-to-image (T2I) generation has garnered significant attention both within the research community and among everyday users. Despite the advancements of T2I models, a common issue encountered by users is the need for repetitive editing of input prompts in order to receive a satisfactory image, which is time-consuming and labor-intensive. Given the demonstrated text generation power of large-scale language models, such as GPT-k, we investigate the potential of utilizing such models to improve the prompt editing process for T2I generation. We conduct a series of experiments to compare the common edits made by humans and GPT-k, evaluate the performance of GPT-k in prompting T2I, and examine factors that may influence this process. We found that GPT-k models focus more on inserting modifiers while humans tend to replace words and phrases, which includes changes to the subject matter. Experimental results show that GPT-k are more effective in adjusting modifiers rather than predicting spontaneous changes in the primary subject matters. Adopting the edit suggested by GPT-k models may reduce the percentage of remaining edits by 20-30%.

Xuehai He, Weixi Feng, Tsu-Jui Fu et al.

Diffusion models, such as Stable Diffusion, have shown incredible performance on text-to-image generation. Since text-to-image generation often requires models to generate visual concepts with fine-grained details and attributes specified in text prompts, can we leverage the powerful representations learned by pre-trained diffusion models for discriminative tasks such as image-text matching? To answer this question, we propose a novel approach, Discriminative Stable Diffusion (DSD), which turns pre-trained text-to-image diffusion models into few-shot discriminative learners. Our approach mainly uses the cross-attention score of a Stable Diffusion model to capture the mutual influence between visual and textual information and fine-tune the model via efficient attention-based prompt learning to perform image-text matching. By comparing DSD with state-of-the-art methods on several benchmark datasets, we demonstrate the potential of using pre-trained diffusion models for discriminative tasks with superior results on few-shot image-text matching.

Tsu-Jui Fu, Linjie Li, Zhe Gan et al.

A great challenge in video-language (VidL) modeling lies in the disconnection between fixed video representations extracted from image/video understanding models and downstream VidL data. Recent studies try to mitigate this disconnection via end-to-end training. To make it computationally feasible, prior works tend to "imagify" video inputs, i.e., a handful of sparsely sampled frames are fed into a 2D CNN, followed by a simple mean-pooling or concatenation to obtain the overall video representations. Although achieving promising results, such simple approaches may lose temporal information that is essential for performing downstream VidL tasks. In this work, we present VIOLET, a fully end-to-end VIdeO-LanguagE Transformer, which adopts a video transformer to explicitly model the temporal dynamics of video inputs. Further, unlike previous studies that found pre-training tasks on video inputs (e.g., masked frame modeling) not very effective, we design a new pre-training task, Masked Visual-token Modeling (MVM), for better video modeling. Specifically, the original video frame patches are "tokenized" into discrete visual tokens, and the goal is to recover the original visual tokens based on the masked patches. Comprehensive analysis demonstrates the effectiveness of both explicit temporal modeling via video transformer and MVM. As a result, VIOLET achieves new state-of-the-art performance on 5 video question answering tasks and 4 text-to-video retrieval tasks.

Tsu-Jui Fu, Xin Eric Wang, William Yang Wang

Despite having promising results, style transfer, which requires preparing style images in advance, may result in lack of creativity and accessibility. Following human instruction, on the other hand, is the most natural way to perform artistic style transfer that can significantly improve controllability for visual effect applications. We introduce a new task, language-driven artistic style transfer (LDAST), to manipulate the style of a content image, guided by a text. We propose contrastive language visual artist (CLVA) that learns to extract visual semantics from style instructions and accomplish LDAST by the patch-wise style discriminator. The discriminator considers the correlation between language and patches of style images or transferred results to jointly embed style instructions. CLVA further compares contrastive pairs of content images and style instructions to improve the mutual relativeness. The results from the same content image can preserve consistent content structures. Besides, they should present analogous style patterns from style instructions that contain similar visual semantics. The experiments show that our CLVA is effective and achieves superb transferred results on LDAST.

Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton et al.

Video editing tools are widely used nowadays for digital design. Although the demand for these tools is high, the prior knowledge required makes it difficult for novices to get started. Systems that could follow natural language instructions to perform automatic editing would significantly improve accessibility. This paper introduces the language-based video editing (LBVE) task, which allows the model to edit, guided by text instruction, a source video into a target video. LBVE contains two features: 1) the scenario of the source video is preserved instead of generating a completely different video; 2) the semantic is presented differently in the target video, and all changes are controlled by the given instruction. We propose a Multi-Modal Multi-Level Transformer (M$^3$L) to carry out LBVE. M$^3$L dynamically learns the correspondence between video perception and language semantic at different levels, which benefits both the video understanding and video frame synthesis. We build three new datasets for evaluation, including two diagnostic and one from natural videos with human-labeled text. Extensive experimental results show that M$^3$L is effective for video editing and that LBVE can lead to a new field toward vision-and-language research.

An Yan, Xin Eric Wang, Tsu-Jui Fu et al.

Recent advances in language and vision push forward the research of captioning a single image to describing visual differences between image pairs. Suppose there are two images, I_1 and I_2, and the task is to generate a description W_{1,2} comparing them, existing methods directly model { I_1, I_2 } -> W_{1,2} mapping without the semantic understanding of individuals. In this paper, we introduce a Learning-to-Compare (L2C) model, which learns to understand the semantic structures of these two images and compare them while learning to describe each one. We demonstrate that L2C benefits from a comparison between explicit semantic representations and single-image captions, and generalizes better on the new testing image pairs. It outperforms the baseline on both automatic evaluation and human evaluation for the Birds-to-Words dataset.

Tsu-Jui Fu, William Yang Wang, Daniel McDuff et al.

Creating presentation materials requires complex multimodal reasoning skills to summarize key concepts and arrange them in a logical and visually pleasing manner. Can machines learn to emulate this laborious process? We present a novel task and approach for document-to-slide generation. Solving this involves document summarization, image and text retrieval, slide structure and layout prediction to arrange key elements in a form suitable for presentation. We propose a hierarchical sequence-to-sequence approach to tackle our task in an end-to-end manner. Our approach exploits the inherent structures within documents and slides and incorporates paraphrasing and layout prediction modules to generate slides. To help accelerate research in this domain, we release a dataset about 6K paired documents and slide decks used in our experiments. We show that our approach outperforms strong baselines and produces slides with rich content and aligned imagery.

Jhih-Wei Chen, Tsu-Jui Fu, Chen-Kang Lee et al.

Although distant supervision automatically generates training data for relation extraction, it also introduces false-positive (FP) and false-negative (FN) training instances to the generated datasets. Whereas both types of errors degrade the final model performance, previous work on distant supervision denoising focuses more on suppressing FP noise and less on resolving the FN problem. We here propose H-FND, a hierarchical false-negative denoising framework for robust distant supervision relation extraction, as an FN denoising solution. H-FND uses a hierarchical policy which first determines whether non-relation (NA) instances should be kept, discarded, or revised during the training process. For those learning instances which are to be revised, the policy further reassigns them appropriate relations, making them better training inputs. Experiments on SemEval-2010 and TACRED were conducted with controlled FN ratios that randomly turn the relations of training and validation instances into negatives to generate FN instances. In this setting, H-FND can revise FN instances correctly and maintains high F1 scores even when 50% of the instances have been turned into negatives. Experiment on NYT10 is further conducted to shows that H-FND is applicable in a realistic setting.

Tsu-Jui Fu, Xin Eric Wang, Scott Grafton et al.

Iterative Language-Based Image Editing (IL-BIE) tasks follow iterative instructions to edit images step by step. Data scarcity is a significant issue for ILBIE as it is challenging to collect large-scale examples of images before and after instruction-based changes. However, humans still accomplish these editing tasks even when presented with an unfamiliar image-instruction pair. Such ability results from counterfactual thinking and the ability to think about alternatives to events that have happened already. In this paper, we introduce a Self-Supervised Counterfactual Reasoning (SSCR) framework that incorporates counterfactual thinking to overcome data scarcity. SSCR allows the model to consider out-of-distribution instructions paired with previous images. With the help of cross-task consistency (CTC), we train these counterfactual instructions in a self-supervised scenario. Extensive results show that SSCR improves the correctness of ILBIE in terms of both object identity and position, establishing a new state of the art (SOTA) on two IBLIE datasets (i-CLEVR and CoDraw). Even with only 50% of the training data, SSCR achieves a comparable result to using complete data.

Wanrong Zhu, Xin Eric Wang, Tsu-Jui Fu et al.

One of the most challenging topics in Natural Language Processing (NLP) is visually-grounded language understanding and reasoning. Outdoor vision-and-language navigation (VLN) is such a task where an agent follows natural language instructions and navigates a real-life urban environment. Due to the lack of human-annotated instructions that illustrate intricate urban scenes, outdoor VLN remains a challenging task to solve. This paper introduces a Multimodal Text Style Transfer (MTST) learning approach and leverages external multimodal resources to mitigate data scarcity in outdoor navigation tasks. We first enrich the navigation data by transferring the style of the instructions generated by Google Maps API, then pre-train the navigator with the augmented external outdoor navigation dataset. Experimental results show that our MTST learning approach is model-agnostic, and our MTST approach significantly outperforms the baseline models on the outdoor VLN task, improving task completion rate by 8.7% relatively on the test set.

Tsu-Jui Fu, Xin Eric Wang, Matthew Peterson et al.

Vision-and-Language Navigation (VLN) is a task where agents must decide how to move through a 3D environment to reach a goal by grounding natural language instructions to the visual surroundings. One of the problems of the VLN task is data scarcity since it is difficult to collect enough navigation paths with human-annotated instructions for interactive environments. In this paper, we explore the use of counterfactual thinking as a human-inspired data augmentation method that results in robust models. Counterfactual thinking is a concept that describes the human propensity to create possible alternatives to life events that have already occurred. We propose an adversarial-driven counterfactual reasoning model that can consider effective conditions instead of low-quality augmented data. In particular, we present a model-agnostic adversarial path sampler (APS) that learns to sample challenging paths that force the navigator to improve based on the navigation performance. APS also serves to do pre-exploration of unseen environments to strengthen the model's ability to generalize. We evaluate the influence of APS on the performance of different VLN baseline models using the room-to-room dataset (R2R). The results show that the adversarial training process with our proposed APS benefits VLN models under both seen and unseen environments. And the pre-exploration process can further gain additional improvements under unseen environments.

Peng-Hsuan Li, Tsu-Jui Fu, Wei-Yun Ma

State-of-the-art approaches of NER have used sequence-labeling BiLSTM as a core module. This paper formally shows the limitation of BiLSTM in modeling cross-context patterns. Two types of simple cross-structures -- self-attention and Cross-BiLSTM -- are shown to effectively remedy the problem. On both OntoNotes 5.0 and WNUT 2017, clear and consistent improvements are achieved over bare-bone models, up to 8.7% on some of the multi-token mentions. In-depth analyses across several aspects of the improvements, especially the identification of multi-token mentions, are further given.

Hsuan-Kung Yang, An-Chieh Cheng, Kuan-Wei Ho et al.

In this paper, we investigate the use of an unsupervised label clustering technique and demonstrate that it enables substantial improvements in visual relationship prediction accuracy on the Person in Context (PIC) dataset. We propose to group object labels with similar patterns of relationship distribution in the dataset into fewer categories. Label clustering not only mitigates both the large classification space and class imbalance issues, but also potentially increases data samples for each clustered category. We further propose to incorporate depth information as an additional feature into the instance segmentation model. The additional depth prediction path supplements the relationship prediction model in a way that bounding boxes or segmentation masks are unable to deliver. We have rigorously evaluated the proposed techniques and performed various ablation analysis to validate the benefits of them.

Zhang-Wei Hong, Tsu-Jui Fu, Tzu-Yun Shann et al.

We present an adversarial active exploration for inverse dynamics model learning, a simple yet effective learning scheme that incentivizes exploration in an environment without any human intervention. Our framework consists of a deep reinforcement learning (DRL) agent and an inverse dynamics model contesting with each other. The former collects training samples for the latter, with an objective to maximize the error of the latter. The latter is trained with samples collected by the former, and generates rewards for the former when it fails to predict the actual action taken by the former. In such a competitive setting, the DRL agent learns to generate samples that the inverse dynamics model fails to predict correctly, while the inverse dynamics model learns to adapt to the challenging samples. We further propose a reward structure that ensures the DRL agent to collect only moderately hard samples but not overly hard ones that prevent the inverse model from predicting effectively. We evaluate the effectiveness of our method on several robotic arm and hand manipulation tasks against multiple baseline models. Experimental results show that our method is comparable to those directly trained with expert demonstrations, and superior to the other baselines even without any human priors.

Yu-Syuan Xu, Tsu-Jui Fu, Hsuan-Kung Yang et al.

In this paper, we present a detailed design of dynamic video segmentation network (DVSNet) for fast and efficient semantic video segmentation. DVSNet consists of two convolutional neural networks: a segmentation network and a flow network. The former generates highly accurate semantic segmentations, but is deeper and slower. The latter is much faster than the former, but its output requires further processing to generate less accurate semantic segmentations. We explore the use of a decision network to adaptively assign different frame regions to different networks based on a metric called expected confidence score. Frame regions with a higher expected confidence score traverse the flow network. Frame regions with a lower expected confidence score have to pass through the segmentation network. We have extensively performed experiments on various configurations of DVSNet, and investigated a number of variants for the proposed decision network. The experimental results show that our DVSNet is able to achieve up to 70.4% mIoU at 19.8 fps on the Cityscape dataset. A high speed version of DVSNet is able to deliver an fps of 30.4 with 63.2% mIoU on the same dataset. DVSNet is also able to reduce up to 95% of the computational workloads.