Diffusion Transformer (DiT) has emerged as the new trend of generative diffusion models on image generation. In view of extremely slow convergence in typical DiT, recent breakthroughs have been driven by mask strategy that significantly improves the training efficiency of DiT with additional intra-image contextual learning. Despite this progress, mask strategy still suffers from two inherent limitations: (a) training-inference discrepancy and (b) fuzzy relations between mask reconstruction & generative diffusion process, resulting in sub-optimal training of DiT. In this work, we address these limitations by novelly unleashing the self-supervised discrimination knowledge to boost DiT training. Technically, we frame our DiT in a teacher-student manner. The teacher-student discriminative pairs are built on the diffusion noises along the same Probability Flow Ordinary Differential Equation (PF-ODE). Instead of applying mask reconstruction loss over both DiT encoder and decoder, we decouple DiT encoder and decoder to separately tackle discriminative and generative objectives. In particular, by encoding discriminative pairs with student and teacher DiT encoders, a new discriminative loss is designed to encourage the inter-image alignment in the self-supervised embedding space. After that, student samples are fed into student DiT decoder to perform the typical generative diffusion task. Extensive experiments are conducted on ImageNet dataset, and our method achieves a competitive balance between training cost and generative capacity.
相關內容
Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.
Text-to-image (T2I) diffusion models have shown significant success in personalized text-to-image generation, which aims to generate novel images with human identities indicated by the reference images. Despite promising identity fidelity has been achieved by several tuning-free methods, they usually suffer from overfitting issues. The learned identity tends to entangle with irrelevant information, resulting in unsatisfied text controllability, especially on faces. In this work, we present MasterWeaver, a test-time tuning-free method designed to generate personalized images with both faithful identity fidelity and flexible editability. Specifically, MasterWeaver adopts an encoder to extract identity features and steers the image generation through additional introduced cross attention. To improve editability while maintaining identity fidelity, we propose an editing direction loss for training, which aligns the editing directions of our MasterWeaver with those of the original T2I model. Additionally, a face-augmented dataset is constructed to facilitate disentangled identity learning, and further improve the editability. Extensive experiments demonstrate that our MasterWeaver can not only generate personalized images with faithful identity, but also exhibit superiority in text controllability. Our code will be publicly available at //github.com/csyxwei/MasterWeaver.
Artificial Intelligence-Generated Content (AIGC) refers to the paradigm of automated content generation utilizing AI models. Mobile AIGC services in the Internet of Vehicles (IoV) network have numerous advantages over traditional cloud-based AIGC services, including enhanced network efficiency, better reconfigurability, and stronger data security and privacy. Nonetheless, AIGC service provisioning frequently demands significant resources. Consequently, resource-constrained roadside units (RSUs) face challenges in maintaining a heterogeneous pool of AIGC services and addressing all user service requests without degrading overall performance. Therefore, in this paper, we propose a decentralized incentive mechanism for mobile AIGC service allocation, employing multi-agent deep reinforcement learning to find the balance between the supply of AIGC services on RSUs and user demand for services within the IoV context, optimizing user experience and minimizing transmission latency. Experimental results demonstrate that our approach achieves superior performance compared to other baseline models.
Text-to-image models are now easy to use and ubiquitous. However, prior work has found that they are prone to recapitulating harmful Western stereotypes. For example, requesting that a model generate an "African person and their house," may produce a person standing next to a straw hut. In this example, the word "African" is an explicit descriptor of the person that the prompt is seeking to depict. Here, we examine whether implicit markers, such as dialect, can also affect the portrayal of people in text-to-image outputs. We pair prompts in Mainstream American English with counterfactuals that express grammatical constructions found in dialects correlated with historically marginalized groups. We find that through minimal, syntax-only changes to prompts, we can systematically shift the skin tone and gender of people in the generated images. We conclude with a discussion of whether dialectic distribution shifts like this are harmful or are expected, possibly even desirable, model behavior.
Information systems (IS) are frequently designed to leverage the negative effect of anchoring bias to influence individuals' decision-making (e.g., by manipulating purchase decisions). Recent advances in Artificial Intelligence (AI) and the explanations of its decisions through explainable AI (XAI) have opened new opportunities for mitigating biased decisions. So far, the potential of these technological advances to overcome anchoring bias remains widely unclear. To this end, we conducted two online experiments with a total of N=390 participants in the context of purchase decisions to examine the impact of AI and XAI-based decision support on anchoring bias. Our results show that AI alone and its combination with XAI help to mitigate the negative effect of anchoring bias. Ultimately, our findings have implications for the design of AI and XAI-based decision support and IS to overcome cognitive biases.
Attention mechanism has been crucial for image diffusion models, however, their quadratic computational complexity limits the sizes of images we can process within reasonable time and memory constraints. This paper investigates the importance of dense attention in generative image models, which often contain redundant features, making them suitable for sparser attention mechanisms. We propose a novel training-free method ToDo that relies on token downsampling of key and value tokens to accelerate Stable Diffusion inference by up to 2x for common sizes and up to 4.5x or more for high resolutions like 2048x2048. We demonstrate that our approach outperforms previous methods in balancing efficient throughput and fidelity.
Diffusion-based text-to-image (T2I) models generate high-fidelity images for given textual prompts. They are trained on large datasets scraped from the Internet, potentially containing unacceptable concepts (e.g., copyright infringing or unsafe). Retraining T2I models after filtering out unacceptable concepts in the training data is inefficient and degrades utility. Hence, there is a need for concept removal techniques (CRTs) which are effective in removing unacceptable concepts, utility-preserving on acceptable concepts, and robust against evasion with adversarial prompts. None of the prior filtering and fine-tuning CRTs satisfy all these requirements simultaneously. We introduce Espresso, the first robust concept filter based on Contrastive Language-Image Pre-Training (CLIP). It identifies unacceptable concepts by projecting the generated image's embedding onto the vector connecting unacceptable and acceptable concepts in the joint text-image embedding space. This ensures robustness by restricting the adversary to adding noise only along this vector, in the direction of the acceptable concept. Further fine-tuning Espresso to separate embeddings of acceptable and unacceptable concepts, while preserving their pairing with image embeddings, ensures both effectiveness and utility. We evaluate Espresso on eleven concepts to show that it is effective (~5% CLIP accuracy on unacceptable concepts), utility-preserving (~93% normalized CLIP score on acceptable concepts), and robust (~4% CLIP accuracy on adversarial prompts for unacceptable concepts). Finally, we present theoretical bounds for the certified robustness of Espresso against adversarial prompts, and an empirical analysis.
Photometric constraint is indispensable for self-supervised monocular depth estimation. It involves warping a source image onto a target view using estimated depth&pose, and then minimizing the difference between the warped and target images. However, the endoscopic built-in light causes significant brightness fluctuations, and thus makes the photometric constraint unreliable. Previous efforts only mitigate this relying on extra models to calibrate image brightness. In this paper, we propose MonoPCC to address the brightness inconsistency radically by reshaping the photometric constraint into a cycle form. Instead of only warping the source image, MonoPCC constructs a closed loop consisting of two opposite forward-backward warping paths: from target to source and then back to target. Thus, the target image finally receives an image cycle-warped from itself, which naturally makes the constraint invariant to brightness changes. Moreover, MonoPCC transplants the source image's phase-frequency into the intermediate warped image to avoid structure lost, and also stabilizes the training via an exponential moving average (EMA) strategy to avoid frequent changes in the forward warping. The comprehensive and extensive experimental results on four endoscopic datasets demonstrate that our proposed MonoPCC shows a great robustness to the brightness inconsistency, and exceeds other state-of-the-arts by reducing the absolute relative error by at least 7.27%, 9.38%, 9.90% and 3.17%, respectively.
Rapid progress in text-to-image generative models coupled with their deployment for visual content creation has magnified the importance of thoroughly evaluating their performance and identifying potential biases. In pursuit of models that generate images that are realistic, diverse, visually appealing, and consistent with the given prompt, researchers and practitioners often turn to automated metrics to facilitate scalable and cost-effective performance profiling. However, commonly-used metrics often fail to account for the full diversity of human preference; often even in-depth human evaluations face challenges with subjectivity, especially as interpretations of evaluation criteria vary across regions and cultures. In this work, we conduct a large, cross-cultural study to study how much annotators in Africa, Europe, and Southeast Asia vary in their perception of geographic representation, visual appeal, and consistency in real and generated images from state-of-the art public APIs. We collect over 65,000 image annotations and 20 survey responses. We contrast human annotations with common automated metrics, finding that human preferences vary notably across geographic location and that current metrics do not fully account for this diversity. For example, annotators in different locations often disagree on whether exaggerated, stereotypical depictions of a region are considered geographically representative. In addition, the utility of automatic evaluations is dependent on assumptions about their set-up, such as the alignment of feature extractors with human perception of object similarity or the definition of "appeal" captured in reference datasets used to ground evaluations. We recommend steps for improved automatic and human evaluations.
Graph Neural Networks (GNNs) are state-of-the-art models for performing prediction tasks on graphs. While existing GNNs have shown great performance on various tasks related to graphs, little attention has been paid to the scenario where out-of-distribution (OOD) nodes exist in the graph during training and inference. Borrowing the concept from CV and NLP, we define OOD nodes as nodes with labels unseen from the training set. Since a lot of networks are automatically constructed by programs, real-world graphs are often noisy and may contain nodes from unknown distributions. In this work, we define the problem of graph learning with out-of-distribution nodes. Specifically, we aim to accomplish two tasks: 1) detect nodes which do not belong to the known distribution and 2) classify the remaining nodes to be one of the known classes. We demonstrate that the connection patterns in graphs are informative for outlier detection, and propose Out-of-Distribution Graph Attention Network (OODGAT), a novel GNN model which explicitly models the interaction between different kinds of nodes and separate inliers from outliers during feature propagation. Extensive experiments show that OODGAT outperforms existing outlier detection methods by a large margin, while being better or comparable in terms of in-distribution classification.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191