Text encoders in diffusion models have rapidly evolved, transitioning from CLIP to T5-XXL. Although this evolution has significantly enhanced the models’ ability to understand complex prompts and generate text, it also leads to a substantial increase in the number of parameters. Despite T5 series encoders being trained on the C4 natural language corpus, which includes a significant amount of non-visual data, diffusion models with T5 encoder do not respond to those non-visual prompts, indicating redundancy in representational power. Therefore, it raises an important question: "Do we really need such a large text encoder?" In pursuit of an answer, we employ vision-based knowledge distillation to train a series of T5 encoder models. To fully inherit T5-XXL’s capabilities, we constructed our dataset based on three criteria: image quality, semantic understanding, and text-rendering. Our results demonstrate the scaling down pattern that the distilled T5-base model can generate images of comparable quality to those produced by T5-XXL, while being 50 times smaller in size. This reduction in model size significantly lowers the GPU requirements for running state-of-the-art models such as FLUX and SD3, making high-quality text-to-image generation more accessible. Our code is available at https: //lifuwang-66.github.io/ScalingDownTE/.

Abstract Vision-and-language navigation (VLN) simulates a visual agent that follows natural-language navigation instructions in real-world scenes. Existing approaches have made enormous progress in navigation in new environments, such as beam search, pre-exploration, and dynamic or hierarchical history encoding. To balance generalization and efficiency, we resort to memorizing visited scenarios apart from the ongoing route while navigating. In this work, we introduce a mechanism of Episodic Scene memory (ESceme) for VLN that wakes an agent’s memories of past visits when it enters the current scene. The episodic scene memory allows the agent to envision a bigger picture of the next prediction. This way, the agent learns to utilize dynamically updated information instead of merely adapting to the current observations. We provide a simple yet effective implementation of ESceme by enhancing the accessible views at each location and progressively completing the memory while navigating. We verify the superiority of ESceme on short-horizon (R2R), long-horizon (R4R), and vision-and-dialog (CVDN) VLN tasks. Our ESceme also wins first place on the CVDN leaderboard. Code is available: https://github.com/qizhust/esceme .

Abstract Aiming to detect novel objects from only a few annotated samples, few-shot object detection (FSOD) has undergone remarkable development. Previous works rarely pay attention to the perspective of gradient propagation to optimize existing methods, therefore failing to make full use of information for novel objects in gradient propagation. We propose a method to solve this problem based on two-stage fine-tuning. A domain adaptation module with multi-constraints is used to promote the spread of gradients, a classification promotion network is used to improve the effect of classification, and a multi-path mask head is added to enrich RoI features. Experiments on PASCAL VOC and COCO datasets show that our model significantly raises the performance compared with previous methods (up to 1–5 $$\%$$ % in average).

Composed image retrieval is a type of image retrieval task where the user provides a reference image as a starting point and specifies a text on how to shift from the starting point to the desired target image. However, most existing methods focus on the composition learning of text and reference images and oversimplify the text as a description, neglecting the inherent structure and the user's shifting intention of the texts. As a result, these methods typically take shortcuts that disregard the visual cue of the reference images. To address this issue, we reconsider the text as instructions and propose a Semantic Shift Network (SSN) that explicitly decomposes the semantic shifts into two steps: from the reference image to the visual prototype and from the visual prototype to the target image. Specifically, SSN explicitly decomposes the instructions into two components: degradation and upgradation, where the degradation is used to picture the visual prototype from the reference image, while the upgradation is used to enrich the visual prototype into the final representations to retrieve the desired target image. The experimental results show that the proposed SSN demonstrates a significant improvement of 5.42% and 1.37% on the CIRR and FashionIQ datasets, respectively, and establishes a new state-of-the-art performance. The code is available at https://github.com/starxing-yuu/SSN.

The recent success of text-to-image generation diffusion models has also revolutionized semantic image editing, enabling the manipulation of images based on query/target texts. Despite these advancements, a significant challenge lies in the potential introduction of contextual prior bias in pre-trained models during image editing, e.g., making unexpected modifications to inappropriate regions. To address this issue, we present a novel approach called Dual-Cycle Diffusion, which generates an unbiased mask to guide image editing. The proposed model incorporates a Bias Elimination Cycle that consists of both a forward path and an inverted path, each featuring a Structural Consistency Cycle to ensure the preservation of image content during the editing process. The forward path utilizes the pre-trained model to produce the edited image, while the inverted path converts the result back to the source image. The unbiased mask is generated by comparing differences between the processed source image and the edited image to ensure that both conform to the same distribution. Our experiments demonstrate the effectiveness of the proposed method, as it significantly improves the D-CLIP score from 0.272 to 0.283. The code will be available athttps://github.com/JohnDreamer/DualCycleDiffsion.

Vision-and-language pre-training models have achieved impressive performance for image captioning. But most of them are trained with millions of paired image-text data and require huge memory and computing overhead. To alleviate this, we try to stand on the shoulders of large-scale pre-trained language models (PLM) and pre-trained vision models (PVM) and efficiently connect them for image captioning. There are two major challenges: one is that language and vision modalities have different semantic granularity (e.g., a noun may cover many pixels), and the other is that the semantic gap still exists between the pre-trained language and vision models. To this end, we design a lightweight and efficient connector to glue PVM and PLM, which holds a criterion of selection-then-transformation . Specifically, in the selection phase, we treat each image as a set of patches instead of pixels. We select salient image patches and cluster them into visual regions to align with text. Then, to effectively reduce the semantic gap, we propose to map the selected image patches into text space through spatial and channel transformations. With training on image captioning datasets, the connector learns to bridge the semantic granularity and semantic gap via backpropagation, preparing for the PLM to generate descriptions. Experimental results on the MSCOCO and Flickr30k datasets demonstrate that our method yields comparable performance to existing works. By solely training the small connector, we achieve a CIDEr performance of 132.2% on the MSCOCO Karpathy test split. Moreover, our findings reveal that fine-tuning the PLM can further enhance performance potential, resulting in a CIDEr score of 140.6%. Code and models are available at https://github.com/YuanEZhou/PrefixCap .

A meaningful video is semantically coherent and changes smoothly. However, most existing fine-grained video representation learning methods learn frame-wise features by aligning frames across videos or exploring relevance between multiple views, neglecting the inherent dynamic process of each video. In this paper, we propose to learn video representations by modeling Video as Stochastic Processes (VSP) via a novel process-based contrastive learning framework, which aims to discriminate between video processes and simultaneously capture the temporal dynamics in the processes. Specifically, we enforce the embeddings of the frame sequence of interest to approximate a goal-oriented stochastic process, i.e., Brownian bridge, in the latent space via a process-based contrastive loss. To construct the Brownian bridge, we adapt specialized sampling strategies under different annotations for both self-supervised and weakly-supervised learning. Experimental results on four datasets show that VSP stands as a state-of-the-art method for various video understanding tasks, including phase progression, phase classification, and frame retrieval. Code is available at https://github.com/hengRUC/VSP.

Paired video and language data is naturally temporal concurrency, which requires the modeling of the temporal dynamics within each modality and the temporal alignment across modalities simultaneously. However, most existing video-language representation learning methods only focus on discrete semantic alignment that encourages aligned semantics to be close in the latent space, or temporal context dependency that captures short-range coherence, failing in building the temporal concurrency. In this paper, we propose to learn video-language representations by modeling video-language pairs as Temporal Concurrent Processes (TCP) via a process-wised distance metric learning framework. Specifically, we employ the soft Dynamic Time Warping (DTW) to measure the distance between two processes across modalities and then optimize the DTW costs. Meanwhile, we further introduce a regularization term that enforces the embeddings of each modality approximating a stochastic process to guarantee the inherent dynamics. Experimental results on three benchmarks demonstrate that TCP stands as a state-of-the-art method for various video-language understanding tasks, including paragraph-to-video retrieval, video moment retrieval, and video question-answering. Code is available at https: //github.com/hengRUC/TCP.

Text-conditional diffusion models are able to generate high-fidelity images with diverse contents. However, linguistic representations frequently exhibit ambiguous descriptions of the envisioned objective imagery, requiring the incorporation of additional control signals to bolster the efficacy of text-guided diffusion models. In this work, we propose Cocktail, a pipeline to mix various modalities into one embedding, amalgamated with a generalized ControlNet (gControlNet), a controllable normalisation (ControlNorm), and a spatial guidance sampling method, to actualize multi-modal and spatially-refined control for text-conditional diffusion models. Specifically, we introduce a hyper-network gControlNet, dedicated to the alignment and infusion of the control signals from disparate modalities into the pre-trained diffusion model. gControlNet is capable of accepting flexible modality signals, encompassing the simultaneous reception of any combination of modality signals, or the supplementary fusion of multiple modality signals. The control signals are then fused and injected into the backbone model according to our proposed ControlNorm. Furthermore, our advanced spatial guidance sampling methodology proficiently incorporates the control signal into the designated region, thereby circumventing the manifestation of undesired objects within the generated image. We demonstrate the results of our method in controlling various modalities, proving high-quality synthesis and fidelity to multiple external signals.

In this work, we explore neat yet effective Transformer-based frameworks for visual grounding. The previous methods generally address the core problem of visual grounding, i.e., multi-modal fusion and reasoning, with manually-designed mechanisms. Such heuristic designs are not only complicated but also make models easily overfit specific data distributions. To avoid this, we first propose TransVG, which establishes multi-modal correspondences by Transformers and localizes referred regions by directly regressing box coordinates. We empirically show that complicated fusion modules can be replaced by a simple stack of Transformer encoder layers with higher performance. However, the core fusion Transformer in TransVG is stand-alone against uni-modal encoders, and thus should be trained from scratch on limited visual grounding data, which makes it hard to be optimized and leads to sub-optimal performance. To this end, we further introduce TransVG++ to make two-fold improvements. For one thing, we upgrade our framework to a purely Transformer-based one by leveraging Vision Transformer (ViT) for vision feature encoding. For another, we devise Language Conditioned Vision Transformer that removes external fusion modules and reuses the uni-modal ViT for vision-language fusion at the intermediate layers. We conduct extensive experiments on five prevalent datasets, and report a series of state-of-the-art records.

We present a method that achieves state-of-the-art results on challenging (few-shot) layout-to-image generation tasks by accurately modeling textures, structures and relationships contained in a complex scene. After compressing RGB images into patch tokens, we propose the Transformer with Focal Attention (TwFA) for exploring dependencies of object-to-object, object-to-patch and patch-to-patch. Compared to existing CNN-based and Transformer-based generation models that entangled modeling on pixel-level&patch-level and object-level&patch-level respectively, the proposed focal attention predicts the current patch token by only focusing on its highly-related tokens that specified by the spatial layout, thereby achieving disambiguation during training. Furthermore, the proposed TwFA largely increases the data efficiency during training, therefore we propose the first few-shot complex scene generation strategy based on the well-trained TwFA. Comprehensive experiments show the superiority of our method, which significantly increases both quantitative metrics and qualitative visual realism with respect to state-of-the-art CNN-based and transformer-based methods. Code is available at https://github.com/JohnDreamer/TwFA.

The training of deep learning-based image deblurring models heavily relies on the paired sharp/blurry image dataset. Although many works verified that synthesized blurry-sharp pairs contribute to improving the deblurring performance, it is still an open problem about how to synthesize realistic and diverse dynamic blurry images. Instead of directly synthesizing blurry images, in this paper, we propose a novel method to generate semantic-aware dense dynamic motion, and employ the generated motion to synthesize blurry images. Specifically, for each sharp image, both the global motion (camera shake) and local motion (object moving) are considered given the depth information as the condition. Then, a blur creation module takes the spatial-variant motion information and the sharp image as input to synthesize a motion-blurred image. A relativistic GAN loss is employed to assure the synthesized blurry image is as realistic as possible. Experiments show that our method can generate diverse dynamic motion and visually realistic blurry images. Also, the generated image pairs can further improve the quantitative performance and generalization ability of the existing deblurring method on several test sets.

Recently, significant progress has been made in masked image modeling to catch up to masked language modeling. However, unlike words in NLP, the lack of semantic decomposition of images still makes masked autoencoding (MAE) different between vision and language. In this paper, we explore a potential visual analogue of words, i.e., semantic parts, and we integrate semantic information into the training process of MAE by proposing a Semantic-Guided Masking strategy. Compared to widely adopted random masking, our masking strategy can gradually guide the network to learn various information, i.e., from intra-part patterns to inter-part relations. In particular, we achieve this in two steps. 1) Semantic part learning: we design a self-supervised part learning method to obtain semantic parts by leveraging and refining the multi-head attention of a ViT-based encoder. 2) Semantic-guided MAE (SemMAE) training: we design a masking strategy that varies from masking a portion of patches in each part to masking a portion of (whole) parts in an image. Extensive experiments on various vision tasks show that SemMAE can learn better image representation by integrating semantic information. In particular, SemMAE achieves 84.5% fine-tuning accuracy on ImageNet-1k, which outperforms the vanilla MAE by 1.4%. In the semantic segmentation and fine-grained recognition tasks, SemMAE also brings significant improvements and yields the state-of-the-art performance.

Grounding natural language in images, such as localizing "the black dog on the left of the tree", is one of the core problems in artificial intelligence, as it needs to comprehend the fine-grained language compositions. However, existing solutions merely rely on the association between the holistic language features and visual features, while neglect the nature of composite reasoning implied in the language. In this paper, we propose a natural language grounding model that can automatically compose a binary tree structure for parsing the language and then perform visual reasoning along the tree in a bottom-up fashion. We call our model RvG-Tree: Recursive Grounding Tree, which is inspired by the intuition that any language expression can be recursively decomposed into two constituent parts, and the grounding confidence score can be recursively accumulated by calculating their grounding scores returned by the two sub-trees.RvG-Tree can be trained end-to-end by using the Straight-Through Gumbel-Softmax estimator that allows the gradients from the continuous score functions passing through the discrete tree construction. Experiments on several benchmarks show that our model achieves the state-of-the-art performance with more explainable reasoning.

With the maturity of visual detection techniques, we are more ambitious in describing visual content with open-vocabulary, fine-grained and free-form language, i.e., the task of image captioning. In particular, we are interested in generating longer, richer and more fine-grained sentences and paragraphs as image descriptions. Image captioning can be translated to the task of sequential language prediction given visual content, where the output sequence forms natural language description with plausible grammar. However, existing image captioning methods focus only on language policy while not visual policy, and thus fail to capture visual context that are crucial for compositional reasoning such as object relationships (e.g., "man riding horse") and visual comparisons (e.g., "small(er) cat"). This issue is especially severe when generating longer sequences such as a paragraph. To fill the gap, we propose a Context-Aware Visual Policy network (CAVP) for fine-grained image-to-language generation: image sentence captioning and image paragraph captioning. During captioning, CAVP explicitly considers the previous visual attentions as context, and decides whether the context is used for the current word/sentence generation given the current visual attention. Compared against traditional visual attention mechanism that only fixes a single visual region at each step, CAVP can attend to complex visual compositions over time. The whole image captioning model-CAVP and its subsequent language policy network-can be efficiently optimized end-to-end by using an actor-critic policy gradient method. We have demonstrated the effectiveness of CAVP by state-of-the-art performances on MS-COCO and Stanford captioning datasets, using various metrics and sensible visualizations of qualitative visual context.

Visual attention not only improves the performance of image captioners, but also serves as a visual interpretation to qualitatively measure the caption rationality and model transparency. Specifically, we expect that a captioner can fix its attentive gaze on the correct objects while generating the corresponding words. This ability is also known as grounded image captioning. However, the grounding accuracy of existing captioners is far from satisfactory.To improve the grounding accuracy while retaining the captioning quality, it is expensive to collect the word-region alignment as strong supervision.To this end, we propose a Part-of-Speech (POS) enhanced image-text matching model (SCAN[24]): POS-SCAN, as the effective knowledge distillation for more grounded image captioning. The benefits are two-fold: 1) given a sentence and an image, POS-SCAN can ground the objects more accurately than SCAN; 2) POS-SCAN serves as a word-region alignment regularization for the captioner's visual attention module. By showing benchmark experimental results, we demonstrate that conventional image captioners equipped with POS-SCAN can significantly improve the grounding accuracy without strong supervision. Last but not the least, we explore the indispensable Self-Critical Sequence Training (SCST[46]) in the context of grounded image captioning and show that the image-text matching score can serve as a reward for more grounded captioning1.

Generating natural language descriptions for videos, i.e., video captioning, essentially requires step-by-step reasoning along the generation process. For example, to generate the sentence “a man is shooting a basketball”, we need to first locate and describe the subject “man”, next reason out the man is “shooting”, then describe the object “basketball” of shooting. However, existing visual reasoning methods designed for visual question answering are not appropriate to video captioning, for it requires more complex visual reasoning on videos over both space and time, and dynamic module composition along the generation process. In this paper, we propose a novel visual reasoning approach for video captioning, named Reasoning Module Networks (RMN), to equip the existing encoder-decoder framework with the above reasoning capacity. Specifically, our RMN employs 1) three sophisticated spatio-temporal reasoning modules, and 2) a dynamic and discrete module selector trained by a linguistic loss with a Gumbel approximation. Extensive experiments on MSVD and MSR-VTT datasets demonstrate the proposed RMN outperforms the state-of-the-art methods while providing an explicit and explainable generation process. Our code is available at https://github.com/tgc1997/RMN.

Visual grounding, a task to ground (i.e., localize) natural language in images, essentially requires composite visual reasoning. However, existing methods over-simplify the composite nature of language into a monolithic sentence embedding or a coarse composition of subject-predicate-object triplet. In this paper, we propose to ground natural language in an intuitive, explainable, and composite fashion as it should be. In particular, we develop a novel modular network called Neural Module Tree network (NMTree) that regularizes the visual grounding along the dependency parsing tree of the sentence, where each node is a neural module that calculates visual attention according to its linguistic feature, and the grounding score is accumulated in a bottom-up direction where as needed. NMTree disentangles the visual grounding from the composite reasoning, allowing the former to only focus on primitive and easy-to-generalize patterns. To reduce the impact of parsing errors, we train the modules and their assembly end-to-end by using the Gumbel-Softmax approximation and its straight-through gradient estimator, accounting for the discrete nature of module assembly. Overall, the proposed NMTree consistently outperforms the state-of-the-arts on several benchmarks. Qualitative results show explainable grounding score calculation in great detail.

Many vision-language tasks can be reduced to the problem of sequence prediction for natural language output. In particular, recent advances in image captioning use deep reinforcement learning (RL) to alleviate the "exposure bias'' during training: ground-truth subsequence is exposed in every step prediction, which introduces bias in test when only predicted subsequence is seen. However, existing RL-based image captioning methods only focus on the language policy while not the visual policy (eg, visual attention), and thus fail to capture the visual context that are crucial for compositional reasoning such as visual relationships (\eg, "man riding horse'') and comparisons (eg. "smaller cat"). To fill the gap, we propose a Context-Aware Visual Policy network (CAVP) for sequence-level image captioning. At every time step, CAVP explicitly accounts for the previous visual attentions as the context, and then decides whether the context is helpful for the current word generation given the current visual attention. Compared against traditional visual attention that only fixes a single image region at every step, CAVP can attend to complex visual compositions over time. The whole image captioning model --- CAVP and its subsequent language policy network --- can be efficiently optimized end-to-end by using an actor-critic policy gradient method with respect to any caption evaluation metric. We demonstrate the effectiveness of CAVP by state-of-the-art performances on MS-COCO offline split and online server, using various metrics and sensible visualizations of qualitative visual context. The code is available at \urlhttps://github.com/daqingliu/CAVP