Abstract:Recently, integrating visual controls into text-to-image~(T2I) models, such as ControlNet method, has received significant attention for finer control capabilities. While various training-free methods make efforts to enhance prompt following in T2I models, the issue with visual control is still rarely studied, especially in the scenario that visual controls are misaligned with text prompts. In this paper, we address the challenge of ``Prompt Following With Visual Control" and propose a training-free approach named Mask-guided Prompt Following (MGPF). Object masks are introduced to distinct aligned and misaligned parts of visual controls and prompts. Meanwhile, a network, dubbed as Masked ControlNet, is designed to utilize these object masks for object generation in the misaligned visual control region. Further, to improve attribute matching, a simple yet efficient loss is designed to align the attention maps of attributes with object regions constrained by ControlNet and object masks. The efficacy and superiority of MGPF are validated through comprehensive quantitative and qualitative experiments.
Abstract:Diffusion models have significantly advanced the state of the art in image, audio, and video generation tasks. However, their applications in practical scenarios are hindered by slow inference speed. Drawing inspiration from the approximation strategies utilized in consistency models, we propose the Sub-path Linear Approximation Model (SLAM), which accelerates diffusion models while maintaining high-quality image generation. SLAM treats the PF-ODE trajectory as a series of PF-ODE sub-paths divided by sampled points, and harnesses sub-path linear (SL) ODEs to form a progressive and continuous error estimation along each individual PF-ODE sub-path. The optimization on such SL-ODEs allows SLAM to construct denoising mappings with smaller cumulative approximated errors. An efficient distillation method is also developed to facilitate the incorporation of more advanced diffusion models, such as latent diffusion models. Our extensive experimental results demonstrate that SLAM achieves an efficient training regimen, requiring only 6 A100 GPU days to produce a high-quality generative model capable of 2 to 4-step generation with high performance. Comprehensive evaluations on LAION, MS COCO 2014, and MS COCO 2017 datasets also illustrate that SLAM surpasses existing acceleration methods in few-step generation tasks, achieving state-of-the-art performance both on FID and the quality of the generated images.
Abstract:Although diffusion models can generate high-quality human images, their applications are limited by the instability in generating hands with correct structures. Some previous works mitigate the problem by considering hand structure yet struggle to maintain style consistency between refined malformed hands and other image regions. In this paper, we aim to solve the problem of inconsistency regarding hand structure and style. We propose a conditional diffusion-based framework RHanDS to refine the hand region with the help of decoupled structure and style guidance. Specifically, the structure guidance is the hand mesh reconstructed from the malformed hand, serving to correct the hand structure. The style guidance is a hand image, e.g., the malformed hand itself, and is employed to furnish the style reference for hand refining. In order to suppress the structure leakage when referencing hand style and effectively utilize hand data to improve the capability of the model, we build a multi-style hand dataset and introduce a twostage training strategy. In the first stage, we use paired hand images for training to generate hands with the same style as the reference. In the second stage, various hand images generated based on the human mesh are used for training to enable the model to gain control over the hand structure. We evaluate our method and counterparts on the test dataset of the proposed multi-style hand dataset. The experimental results show that RHanDS can effectively refine hands structure- and style- correctly compared with previous methods. The codes and datasets will be available soon.
Abstract:In Conversational Recommendation Systems (CRS), the central question is how the conversational agent can naturally ask for user preferences and provide suitable recommendations. Existing works mainly follow the hierarchical architecture, where a higher policy decides whether to invoke the conversation module (to ask questions) or the recommendation module (to make recommendations). This architecture prevents these two components from fully interacting with each other. In contrast, this paper proposes a novel architecture, the long short-term feedback architecture, to connect these two essential components in CRS. Specifically, the recommendation predicts the long-term recommendation target based on the conversational context and the user history. Driven by the targeted recommendation, the conversational model predicts the next topic or attribute to verify if the user preference matches the target. The balance feedback loop continues until the short-term planner output matches the long-term planner output, that is when the system should make the recommendation.
Abstract:Recently, Multi-Scenario Learning (MSL) is widely used in recommendation and retrieval systems in the industry because it facilitates transfer learning from different scenarios, mitigating data sparsity and reducing maintenance cost. These efforts produce different MSL paradigms by searching more optimal network structure, such as Auxiliary Network, Expert Network, and Multi-Tower Network. It is intuitive that different scenarios could hold their specific characteristics, activating the user's intents quite differently. In other words, different kinds of auxiliary features would bear varying importance under different scenarios. With more discriminative feature representations refined in a scenario-aware manner, better ranking performance could be easily obtained without expensive search for the optimal network structure. Unfortunately, this simple idea is mainly overlooked but much desired in real-world systems.Further analysis also validates the rationality of adaptive feature learning under a multi-scenario scheme. Moreover, our A/B test results on the Alibaba search advertising platform also demonstrate that Maria is superior in production environments.
Abstract:Extracting expressive visual features is crucial for accurate Click-Through-Rate (CTR) prediction in visual search advertising systems. Current commercial systems use off-the-shelf visual encoders to facilitate fast online service. However, the extracted visual features are coarse-grained and/or biased. In this paper, we present a visual encoding framework for CTR prediction to overcome these problems. The framework is based on contrastive learning which pulls positive pairs closer and pushes negative pairs apart in the visual feature space. To obtain fine-grained visual features,we present contrastive learning supervised by click through data to fine-tune the visual encoder. To reduce sample selection bias, firstly we train the visual encoder offline by leveraging both unbiased self-supervision and click supervision signals. Secondly, we incorporate a debiasing network in the online CTR predictor to adjust the visual features by contrasting high impression items with selected items with lower impressions.We deploy the framework in the visual sponsor search system at Alibaba. Offline experiments on billion-scale datasets and online experiments demonstrate that the proposed framework can make accurate and unbiased predictions.
Abstract:Multimodal supervision has achieved promising results in many visual language understanding tasks, where the language plays an essential role as a hint or context for recognizing and locating instances. However, due to the defects of the human-annotated language corpus, multimodal supervision remains unexplored in fully supervised object detection scenarios. In this paper, we take advantage of language prompt to introduce effective and unbiased linguistic supervision into object detection, and propose a new mechanism called multimodal knowledge learning (\textbf{MKL}), which is required to learn knowledge from language supervision. Specifically, we design prompts and fill them with the bounding box annotations to generate descriptions containing extensive hints and context for instances recognition and localization. The knowledge from language is then distilled into the detection model via maximizing cross-modal mutual information in both image- and object-level. Moreover, the generated descriptions are manipulated to produce hard negatives to further boost the detector performance. Extensive experiments demonstrate that the proposed method yields a consistent performance gain by 1.6\% $\sim$ 2.1\% and achieves state-of-the-art on MS-COCO and OpenImages datasets.
Abstract:This report demonstrates our solution for the Open Images 2018 Challenge. Based on our detailed analysis on the Open Images Datasets (OID), it is found that there are four typical features: large-scale, hierarchical tag system, severe annotation incompleteness and data imbalance. Considering these characteristics, an amount of strategies are employed, including SNIPER, soft sampling, class-aware sampling (CAS), hierarchical non-maximum suppression (HNMS) and so on. In virtue of these effective strategies, and further using the powerful SENet154 armed with feature pyramid module and deformable ROIalign as the backbone, our best single model could achieve a mAP of 56.9%. After a further ensemble with 9 models, the final mAP is boosted to 62.2% in the public leaderboard (ranked the 2nd place) and 58.6% in the private leaderboard (ranked the 3rd place, slightly inferior to the 1st place by only 0.04 point).