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.
相關內容
This paper investigates a 2D to 3D image translation method with a straightforward technique, enabling correlated 2D X-ray to 3D CT-like reconstruction. We observe that existing approaches, which integrate information across multiple 2D views in the latent space, lose valuable signal information during latent encoding. Instead, we simply repeat and concatenate the 2D views into higher-channel 3D volumes and approach the 3D reconstruction challenge as a straightforward 3D to 3D generative modeling problem, sidestepping several complex modeling issues. This method enables the reconstructed 3D volume to retain valuable information from the 2D inputs, which are passed between channel states in a Swin UNETR backbone. Our approach applies neural optimal transport, which is fast and stable to train, effectively integrating signal information across multiple views without the requirement for precise alignment; it produces non-collapsed reconstructions that are highly faithful to the 2D views, even after limited training. We demonstrate correlated results, both qualitatively and quantitatively, having trained our model on a single dataset and evaluated its generalization ability across six datasets, including out-of-distribution samples.
The 3D Gaussian Splatting technique has significantly advanced the construction of radiance fields from multi-view images, enabling real-time rendering. While point-based rasterization effectively reduces computational demands for rendering, it often struggles to accurately reconstruct the geometry of the target object, especially under strong lighting. To address this challenge, we introduce a novel approach that combines octree-based implicit surface representations with Gaussian splatting. Our method consists of four stages. Initially, it reconstructs a signed distance field (SDF) and a radiance field through volume rendering, encoding them in a low-resolution octree. The initial SDF represents the coarse geometry of the target object. Subsequently, it introduces 3D Gaussians as additional degrees of freedom, which are guided by the SDF. In the third stage, the optimized Gaussians further improve the accuracy of the SDF, allowing it to recover finer geometric details compared to the initial SDF obtained in the first stage. Finally, it adopts the refined SDF to further optimize the 3D Gaussians via splatting, eliminating those that contribute little to visual appearance. Experimental results show that our method, which leverages the distribution of 3D Gaussians with SDFs, reconstructs more accurate geometry, particularly in images with specular highlights caused by strong lighting.
Text-based 2D diffusion models have demonstrated impressive capabilities in image generation and editing. Meanwhile, the 2D diffusion models also exhibit substantial potentials for 3D editing tasks. However, how to achieve consistent edits across multiple viewpoints remains a challenge. While the iterative dataset update method is capable of achieving global consistency, it suffers from slow convergence and over-smoothed textures. We propose SyncNoise, a novel geometry-guided multi-view consistent noise editing approach for high-fidelity 3D scene editing. SyncNoise synchronously edits multiple views with 2D diffusion models while enforcing multi-view noise predictions to be geometrically consistent, which ensures global consistency in both semantic structure and low-frequency appearance. To further enhance local consistency in high-frequency details, we set a group of anchor views and propagate them to their neighboring frames through cross-view reprojection. To improve the reliability of multi-view correspondences, we introduce depth supervision during training to enhance the reconstruction of precise geometries. Our method achieves high-quality 3D editing results respecting the textual instructions, especially in scenes with complex textures, by enhancing geometric consistency at the noise and pixel levels.
Customized image generation, which seeks to synthesize images with consistent characters, holds significant relevance for applications such as storytelling, portrait generation, and character design. However, previous approaches have encountered challenges in preserving characters with high-fidelity consistency due to inadequate feature extraction and concept confusion of reference characters. Therefore, we propose Character-Adapter, a plug-and-play framework designed to generate images that preserve the details of reference characters, ensuring high-fidelity consistency. Character-Adapter employs prompt-guided segmentation to ensure fine-grained regional features of reference characters and dynamic region-level adapters to mitigate concept confusion. Extensive experiments are conducted to validate the effectiveness of Character-Adapter. Both quantitative and qualitative results demonstrate that Character-Adapter achieves the state-of-the-art performance of consistent character generation, with an improvement of 24.8% compared with other methods
Matching raw audio signals with textual descriptions requires understanding the audio's content and the description's semantics and then drawing connections between the two modalities. This paper investigates a hybrid retrieval system that utilizes audio metadata as an additional clue to understand the content of audio signals before matching them with textual queries. We experimented with metadata often attached to audio recordings, such as keywords and natural-language descriptions, and we investigated late and mid-level fusion strategies to merge audio and metadata. Our hybrid approach with keyword metadata and late fusion improved the retrieval performance over a content-based baseline by 2.36 and 3.69 pp. mAP@10 on the ClothoV2 and AudioCaps benchmarks, respectively.
Sparse-view computed tomography (SVCT) reconstruction aims to acquire CT images based on sparsely-sampled measurements. It allows the subjects exposed to less ionizing radiation, reducing the lifetime risk of developing cancers. Recent researches employ implicit neural representation (INR) techniques to reconstruct CT images from a single SV sinogram. However, due to ill-posedness, these INR-based methods may leave considerable ``holes'' (i.e., unmodeled spaces) in their fields, leading to sub-optimal results. In this paper, we propose the Coordinate-based Continuous Projection Field (CoCPF), which aims to build hole-free representation fields for SVCT reconstruction, achieving better reconstruction quality. Specifically, to fill the holes, CoCPF first employs the stripe-based volume sampling module to broaden the sampling regions of Radon transformation from rays (1D space) to stripes (2D space), which can well cover the internal regions between SV projections. Then, by feeding the sampling regions into the proposed differentiable rendering modules, the holes can be jointly optimized during training, reducing the ill-posed levels. As a result, CoCPF can accurately estimate the internal measurements between SV projections (i.e., DV sinograms), producing high-quality CT images after re-projection. Extensive experiments on simulated and real projection datasets demonstrate that CoCPF outperforms state-of-the-art methods for 2D and 3D SVCT reconstructions under various projection numbers and geometries, yielding fine-grained details and fewer artifacts. Our code will be publicly available.
Most medical image lesion segmentation methods rely on hand-crafted accurate annotations of the original image for supervised learning. Recently, a series of weakly supervised or unsupervised methods have been proposed to reduce the dependence on pixel-level annotations. However, these methods are essentially based on pixel-level annotation, ignoring the image-level diagnostic results of the current massive medical images. In this paper, we propose a dual U-shaped two-stage framework that utilizes image-level labels to prompt the segmentation. In the first stage, we pre-train a classification network with image-level labels, which is used to obtain the hierarchical pyramid features and guide the learning of downstream branches. In the second stage, we feed the hierarchical features obtained from the classification branch into the downstream branch through short-skip and long-skip and get the lesion masks under the supervised learning of pixel-level labels. Experiments show that our framework achieves better results than networks simply using pixel-level annotations.
While text-to-visual models now produce photo-realistic images and videos, they struggle with compositional text prompts involving attributes, relationships, and higher-order reasoning such as logic and comparison. In this work, we conduct an extensive human study on GenAI-Bench to evaluate the performance of leading image and video generation models in various aspects of compositional text-to-visual generation. We also compare automated evaluation metrics against our collected human ratings and find that VQAScore -- a metric measuring the likelihood that a VQA model views an image as accurately depicting the prompt -- significantly outperforms previous metrics such as CLIPScore. In addition, VQAScore can improve generation in a black-box manner (without finetuning) via simply ranking a few (3 to 9) candidate images. Ranking by VQAScore is 2x to 3x more effective than other scoring methods like PickScore, HPSv2, and ImageReward at improving human alignment ratings for DALL-E 3 and Stable Diffusion, especially on compositional prompts that require advanced visio-linguistic reasoning. We will release a new GenAI-Rank benchmark with over 40,000 human ratings to evaluate scoring metrics on ranking images generated from the same prompt. Lastly, we discuss promising areas for improvement in VQAScore, such as addressing fine-grained visual details. We will release all human ratings (over 80,000) to facilitate scientific benchmarking of both generative models and automated metrics.
Accurately estimating depth in 360-degree imagery is crucial for virtual reality, autonomous navigation, and immersive media applications. Existing depth estimation methods designed for perspective-view imagery fail when applied to 360-degree images due to different camera projections and distortions, whereas 360-degree methods perform inferior due to the lack of labeled data pairs. We propose a new depth estimation framework that utilizes unlabeled 360-degree data effectively. Our approach uses state-of-the-art perspective depth estimation models as teacher models to generate pseudo labels through a six-face cube projection technique, enabling efficient labeling of depth in 360-degree images. This method leverages the increasing availability of large datasets. Our approach includes two main stages: offline mask generation for invalid regions and an online semi-supervised joint training regime. We tested our approach on benchmark datasets such as Matterport3D and Stanford2D3D, showing significant improvements in depth estimation accuracy, particularly in zero-shot scenarios. Our proposed training pipeline can enhance any 360 monocular depth estimator and demonstrates effective knowledge transfer across different camera projections and data types. See our project page for results: //albert100121.github.io/Depth-Anywhere/
We propose PLAYER*, a novel framework that addresses the limitations of existing agent-based approaches built on Large Language Models (LLMs) in handling complex questions and understanding interpersonal relationships in dynamic environments. PLAYER* enhances path planning in Murder Mystery Games (MMGs) using an anytime sampling-based planner and a questioning-driven search framework. By equipping agents with a set of sensors, PLAYER* eliminates the need for pre-defined questions and enables agents to navigate complex social interactions. We additionally make a contribution by introducing a quantifiable evaluation method using multiple-choice questions and present WellPlay, a dataset containing 1,482 question-answer pairs. Experimental results demonstrate PLAYER*'s superiority over existing multi-agent methods, enhancing the generalisability and adaptability of agents in MMGs and paving the way for more effective multi-agent interactions.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191