Cross-Attention Makes Inference Cumbersome in Text-to-Image Diffusion Models

This study explores the role of cross-attention during inference in text-conditional diffusion models. We find that cross-attention outputs converge to a fixed point after few inference steps. Accordingly, the time point of convergence naturally divides the entire inference process into two stages: an initial semantics-planning stage, during which, the model relies on cross-attention to plan text-oriented visual semantics, and a subsequent fidelity-improving stage, during which the model tries to generate images from previously planned semantics. Surprisingly, ignoring text conditions in the fidelity-improving stage not only reduces computation complexity, but also maintains model performance. This yields a simple and training-free method called TGATE for efficient generation, which caches the cross-attention output once it converges and keeps it fixed during the remaining inference steps. Our empirical study on the MS-COCO validation set confirms its effectiveness. The source code of TGATE is available at https://github.com/HaozheLiu-ST/T-GATE.

Diffusion-Driven Self-Supervised Learning for Shape Reconstruction and Pose Estimation

Fully-supervised category-level pose estimation aims to determine the 6-DoF poses of unseen instances from known categories, requiring expensive mannual labeling costs. Recently, various self-supervised category-level pose estimation methods have been proposed to reduce the requirement of the annotated datasets. However, most methods rely on synthetic data or 3D CAD model for self-supervised training, and they are typically limited to addressing single-object pose problems without considering multi-objective tasks or shape reconstruction. To overcome these challenges and limitations, we introduce a diffusion-driven self-supervised network for multi-object shape reconstruction and categorical pose estimation, only leveraging the shape priors. Specifically, to capture the SE(3)-equivariant pose features and 3D scale-invariant shape information, we present a Prior-Aware Pyramid 3D Point Transformer in our network. This module adopts a point convolutional layer with radial-kernels for pose-aware learning and a 3D scale-invariant graph convolution layer for object-level shape representation, respectively. Furthermore, we introduce a pretrain-to-refine self-supervised training paradigm to train our network. It enables proposed network to capture the associations between shape priors and observations, addressing the challenge of intra-class shape variations by utilising the diffusion mechanism. Extensive experiments conducted on four public datasets and a self-built dataset demonstrate that our method significantly outperforms state-of-the-art self-supervised category-level baselines and even surpasses some fully-supervised instance-level and category-level methods.

DragAnything: Motion Control for Anything using Entity Representation

We introduce DragAnything, which utilizes a entity representation to achieve motion control for any object in controllable video generation. Comparison to existing motion control methods, DragAnything offers several advantages. Firstly, trajectory-based is more userfriendly for interaction, when acquiring other guidance signals (e.g., masks, depth maps) is labor-intensive. Users only need to draw a line (trajectory) during interaction. Secondly, our entity representation serves as an open-domain embedding capable of representing any object, enabling the control of motion for diverse entities, including background. Lastly, our entity representation allows simultaneous and distinct motion control for multiple objects. Extensive experiments demonstrate that our DragAnything achieves state-of-the-art performance for FVD, FID, and User Study, particularly in terms of object motion control, where our method surpasses the previous methods (e.g., DragNUWA) by 26% in human voting.

Bring Your Own Character: A Holistic Solution for Automatic Facial Animation Generation of Customized Characters

Animating virtual characters has always been a fundamental research problem in virtual reality (VR). Facial animations play a crucial role as they effectively convey emotions and attitudes of virtual humans. However, creating such facial animations can be challenging, as current methods often involve utilization of expensive motion capture devices or significant investments of time and effort from human animators in tuning animation parameters. In this paper, we propose a holistic solution to automatically animate virtual human faces. In our solution, a deep learning model was first trained to retarget the facial expression from input face images to virtual human faces by estimating the blendshape coefficients. This method offers the flexibility of generating animations with characters of different appearances and blendshape topologies. Second, a practical toolkit was developed using Unity 3D, making it compatible with the most popular VR applications. The toolkit accepts both image and video as input to animate the target virtual human faces and enables users to manipulate the animation results. Furthermore, inspired by the spirit of Human-in-the-loop (HITL), we leveraged user feedback to further improve the performance of the model and toolkit, thereby increasing the customization properties to suit user preferences. The whole solution, for which we will make the code public, has the potential to accelerate the generation of facial animations for use in VR applications.

Skip \n: A Simple Method to Reduce Hallucination in Large Vision-Language Models

Recent advancements in large vision-language models (LVLMs) have demonstrated impressive capability in visual information understanding with human language. Despite these advances, LVLMs still face challenges with multimodal hallucination, such as generating text descriptions of objects that are not present in the visual information. However, the underlying fundamental reasons of multimodal hallucinations remain poorly explored. In this paper, we propose a new perspective, suggesting that the inherent biases in LVLMs might be a key factor in hallucinations. Specifically, we systematically identify a semantic shift bias related to paragraph breaks (\n\n), where the content before and after '\n\n' in the training data frequently exhibit significant semantic changes. This pattern leads the model to infer that the contents following '\n\n' should be obviously different from the preceding contents with less hallucinatory descriptions, thereby increasing the probability of hallucinatory descriptions subsequent to the '\n\n'. We have validated this hypothesis on multiple publicly available LVLMs. Besides, we find that deliberately inserting '\n\n' at the generated description can induce more hallucinations. A simple method is proposed to effectively mitigate the hallucination of LVLMs by skipping the output of '\n'.

Skip $\textbackslash n$: A simple method to reduce hallucination in Large Vision-Language Models

Recent advancements in large vision-language models (LVLMs) have demonstrated impressive capability in visual information understanding with human language. Despite these advances, LVLMs still face challenges with multimodal hallucination, such as generating text descriptions of objects that are not present in the visual information. However, the underlying fundamental reasons of multimodal hallucinations remain poorly explored. In this paper, we propose a new perspective, suggesting that the inherent biases in LVLMs might be a key factor in hallucinations. Specifically, we systematically identify a semantic shift bias related to paragraph breaks ('$\textbackslash n\textbackslash n$'), where the content before and after '$\textbackslash n\textbackslash n$' in the training data frequently exhibit significant semantic changes. This pattern leads the model to infer that the contents following '$\textbackslash n\textbackslash n$' should be obviously different from the preceding contents with less hallucinatory descriptions, thereby increasing the probability of hallucinatory descriptions subsequent to the '$\textbackslash n\textbackslash n$'. We have validated this hypothesis on multiple publicly available LVLMs. Besides, we find that deliberately inserting '$\textbackslash n\textbackslash n$' at the generated description can induce more hallucinations. A simple method is proposed to effectively mitigate the hallucination of LVLMs by skipping the output of `\textbackslash n'.

Delocate: Detection and Localization for Deepfake Videos with Randomly-Located Tampered Traces

Deepfake videos are becoming increasingly realistic, showing subtle tampering traces on facial areasthat vary between frames. Consequently, many existing Deepfake detection methods struggle to detect unknown domain Deepfake videos while accurately locating the tampered region. To address thislimitation, we propose Delocate, a novel Deepfake detection model that can both recognize andlocalize unknown domain Deepfake videos. Ourmethod consists of two stages named recoveringand localization. In the recovering stage, the modelrandomly masks regions of interest (ROIs) and reconstructs real faces without tampering traces, resulting in a relatively good recovery effect for realfaces and a poor recovery effect for fake faces. Inthe localization stage, the output of the recoveryphase and the forgery ground truth mask serve assupervision to guide the forgery localization process. This process strategically emphasizes the recovery phase of fake faces with poor recovery, facilitating the localization of tampered regions. Ourextensive experiments on four widely used benchmark datasets demonstrate that Delocate not onlyexcels in localizing tampered areas but also enhances cross-domain detection performance.

Towards A Better Metric for Text-to-Video Generation

Generative models have demonstrated remarkable capability in synthesizing high-quality text, images, and videos. For video generation, contemporary text-to-video models exhibit impressive capabilities, crafting visually stunning videos. Nonetheless, evaluating such videos poses significant challenges. Current research predominantly employs automated metrics such as FVD, IS, and CLIP Score. However, these metrics provide an incomplete analysis, particularly in the temporal assessment of video content, thus rendering them unreliable indicators of true video quality. Furthermore, while user studies have the potential to reflect human perception accurately, they are hampered by their time-intensive and laborious nature, with outcomes that are often tainted by subjective bias. In this paper, we investigate the limitations inherent in existing metrics and introduce a novel evaluation pipeline, the Text-to-Video Score (T2VScore). This metric integrates two pivotal criteria: (1) Text-Video Alignment, which scrutinizes the fidelity of the video in representing the given text description, and (2) Video Quality, which evaluates the video's overall production caliber with a mixture of experts. Moreover, to evaluate the proposed metrics and facilitate future improvements on them, we present the TVGE dataset, collecting human judgements of 2,543 text-to-video generated videos on the two criteria. Experiments on the TVGE dataset demonstrate the superiority of the proposed T2VScore on offering a better metric for text-to-video generation.

Moonshot: Towards Controllable Video Generation and Editing with Multimodal Conditions

Most existing video diffusion models (VDMs) are limited to mere text conditions. Thereby, they are usually lacking in control over visual appearance and geometry structure of the generated videos. This work presents Moonshot, a new video generation model that conditions simultaneously on multimodal inputs of image and text. The model builts upon a core module, called multimodal video block (MVB), which consists of conventional spatialtemporal layers for representing video features, and a decoupled cross-attention layer to address image and text inputs for appearance conditioning. In addition, we carefully design the model architecture such that it can optionally integrate with pre-trained image ControlNet modules for geometry visual conditions, without needing of extra training overhead as opposed to prior methods. Experiments show that with versatile multimodal conditioning mechanisms, Moonshot demonstrates significant improvement on visual quality and temporal consistency compared to existing models. In addition, the model can be easily repurposed for a variety of generative applications, such as personalized video generation, image animation and video editing, unveiling its potential to serve as a fundamental architecture for controllable video generation. Models will be made public on https://github.com/salesforce/LAVIS.