High-quality driving video generation is crucial for providing training data for autonomous driving models. However, current generative models rarely focus on enhancing camera motion control under multi-view tasks, which is essential for driving video generation. Therefore, we propose MyGo, an end-to-end framework for video generation, introducing motion of onboard cameras as conditions to make progress in camera controllability and multi-view consistency. MyGo employs additional plug-in modules to inject camera parameters into the pre-trained video diffusion model, which retains the extensive knowledge of the pre-trained model as much as possible. Furthermore, we use epipolar constraints and neighbor view information during the generation process of each view to enhance spatial-temporal consistency. Experimental results show that MyGo has achieved state-of-the-art results in both general camera-controlled video generation and multi-view driving video generation tasks, which lays the foundation for more accurate environment simulation in autonomous driving. Project page: https://metadrivescape.github.io/papers_project/MyGo/page.html
In recent years, the development of diffusion models has led to significant progress in image and video generation tasks, with pre-trained models like the Stable Diffusion series playing a crucial role. Inspired by model pruning which lightens large pre-trained models by removing unimportant parameters, we propose a novel model fine-tuning method to make full use of these ineffective parameters and enable the pre-trained model with new task-specified capabilities. In this work, we first investigate the importance of parameters in pre-trained diffusion models, and discover that the smallest 10% to 20% of parameters by absolute values do not contribute to the generation process. Based on this observation, we propose a method termed SaRA that re-utilizes these temporarily ineffective parameters, equating to optimizing a sparse weight matrix to learn the task-specific knowledge. To mitigate overfitting, we propose a nuclear-norm-based low-rank sparse training scheme for efficient fine-tuning. Furthermore, we design a new progressive parameter adjustment strategy to make full use of the re-trained/finetuned parameters. Finally, we propose a novel unstructural backpropagation strategy, which significantly reduces memory costs during fine-tuning. Our method enhances the generative capabilities of pre-trained models in downstream applications and outperforms traditional fine-tuning methods like LoRA in maintaining model's generalization ability. We validate our approach through fine-tuning experiments on SD models, demonstrating significant improvements. SaRA also offers a practical advantage that requires only a single line of code modification for efficient implementation and is seamlessly compatible with existing methods.
Autoregressive transformers have revolutionized generative models in language processing and shown substantial promise in image and video generation. However, these models face significant challenges when extended to 3D generation tasks due to their reliance on next-token prediction to learn token sequences, which is incompatible with the unordered nature of 3D data. Instead of imposing an artificial order on 3D data, in this paper, we introduce G3PT, a scalable coarse-to-fine 3D generative model utilizing a cross-scale querying transformer. The key is to map point-based 3D data into discrete tokens with different levels of detail, naturally establishing a sequential relationship between different levels suitable for autoregressive modeling. Additionally, the cross-scale querying transformer connects tokens globally across different levels of detail without requiring an ordered sequence. Benefiting from this approach, G3PT features a versatile 3D generation pipeline that effortlessly supports diverse conditional structures, enabling the generation of 3D shapes from various types of conditions. Extensive experiments demonstrate that G3PT achieves superior generation quality and generalization ability compared to previous 3D generation methods. Most importantly, for the first time in 3D generation, scaling up G3PT reveals distinct power-law scaling behaviors.
Recent advancements in generative models have provided promising solutions for synthesizing realistic driving videos, which are crucial for training autonomous driving perception models. However, existing approaches often struggle with multi-view video generation due to the challenges of integrating 3D information while maintaining spatial-temporal consistency and effectively learning from a unified model. We propose DriveScape, an end-to-end framework for multi-view, 3D condition-guided video generation, capable of producing 1024 x 576 high-resolution videos at 10Hz. Unlike other methods limited to 2Hz due to the 3D box annotation frame rate, DriveScape overcomes this with its ability to operate under sparse conditions. Our Bi-Directional Modulated Transformer (BiMot) ensures precise alignment of 3D structural information, maintaining spatial-temporal consistency. DriveScape excels in video generation performance, achieving state-of-the-art results on the nuScenes dataset with an FID score of 8.34 and an FVD score of 76.39. Our project homepage: https://metadrivescape.github.io/papers_project/drivescapev1/index.html
The global self-attention mechanism in diffusion transformers involves redundant computation due to the sparse and redundant nature of visual information, and the attention map of tokens within a spatial window shows significant similarity. To address this redundancy, we propose the Proxy-Tokenized Diffusion Transformer (PT-DiT), which employs sparse representative token attention (where the number of representative tokens is much smaller than the total number of tokens) to model global visual information efficiently. Specifically, within each transformer block, we compute an averaging token from each spatial-temporal window to serve as a proxy token for that region. The global semantics are captured through the self-attention of these proxy tokens and then injected into all latent tokens via cross-attention. Simultaneously, we introduce window and shift window attention to address the limitations in detail modeling caused by the sparse attention mechanism. Building on the well-designed PT-DiT, we further develop the Qihoo-T2X family, which includes a variety of models for T2I, T2V, and T2MV tasks. Experimental results show that PT-DiT achieves competitive performance while reducing the computational complexity in both image and video generation tasks (e.g., a 49% reduction compared to DiT and a 34% reduction compared to PixArt-$\alpha$). The visual exhibition and source code of Qihoo-T2X is available at https://360cvgroup.github.io/Qihoo-T2X/.
Recent advances in diffusion models have improved controllable streetscape generation and supported downstream perception and planning tasks. However, challenges remain in accurately modeling driving scenes and generating long videos. To alleviate these issues, we propose DreamForge, an advanced diffusion-based autoregressive video generation model tailored for 3D-controllable long-term generation. To enhance the lane and foreground generation, we introduce perspective guidance and integrate object-wise position encoding to incorporate local 3D correlation and improve foreground object modeling. We also propose motion-aware temporal attention to capture motion cues and appearance changes in videos. By leveraging motion frames and an autoregressive generation paradigm, we can autoregressively generate long videos (over 200 frames) using a 7-frame model, achieving superior quality compared to the baseline in 16-frame video evaluations. Finally, we integrate our method with the realistic simulation platform DriveArena to provide more reliable open-loop and closed-loop evaluations for vision-based driving agents. The project page is available at https://pjlab-adg.github.io/DriveArena/dreamforge.
Significant progress has been made in text-to-video generation through the use of powerful generative models and large-scale internet data. However, substantial challenges remain in precisely controlling individual concepts within the generated video, such as the motion and appearance of specific characters and the movement of viewpoints. In this work, we propose a novel paradigm that generates each concept in 3D representation separately and then composes them with priors from Large Language Models (LLM) and 2D diffusion models. Specifically, given an input textual prompt, our scheme consists of three stages: 1) We leverage LLM as the director to first decompose the complex query into several sub-prompts that indicate individual concepts within the video~(\textit{e.g.}, scene, objects, motions), then we let LLM to invoke pre-trained expert models to obtain corresponding 3D representations of concepts. 2) To compose these representations, we prompt multi-modal LLM to produce coarse guidance on the scales and coordinates of trajectories for the objects. 3) To make the generated frames adhere to natural image distribution, we further leverage 2D diffusion priors and use Score Distillation Sampling to refine the composition. Extensive experiments demonstrate that our method can generate high-fidelity videos from text with diverse motion and flexible control over each concept. Project page: \url{https://aka.ms/c3v}.
Generating high-fidelity, temporally consistent videos in autonomous driving scenarios faces a significant challenge, e.g. problematic maneuvers in corner cases. Despite recent video generation works are proposed to tackcle the mentioned problem, i.e. models built on top of Diffusion Transformers (DiT), works are still missing which are targeted on exploring the potential for multi-view videos generation scenarios. Noticeably, we propose the first DiT-based framework specifically designed for generating temporally and multi-view consistent videos which precisely match the given bird's-eye view layouts control. Specifically, the proposed framework leverages a parameter-free spatial view-inflated attention mechanism to guarantee the cross-view consistency, where joint cross-attention modules and ControlNet-Transformer are integrated to further improve the precision of control. To demonstrate our advantages, we extensively investigate the qualitative comparisons on nuScenes dataset, particularly in some most challenging corner cases. In summary, the effectiveness of our proposed method in producing long, controllable, and highly consistent videos under difficult conditions is proven to be effective.
Estimating video depth in open-world scenarios is challenging due to the diversity of videos in appearance, content motion, camera movement, and length. We present DepthCrafter, an innovative method for generating temporally consistent long depth sequences with intricate details for open-world videos, without requiring any supplementary information such as camera poses or optical flow. The generalization ability to open-world videos is achieved by training the video-to-depth model from a pre-trained image-to-video diffusion model, through our meticulously designed three-stage training strategy. Our training approach enables the model to generate depth sequences with variable lengths at one time, up to 110 frames, and harvest both precise depth details and rich content diversity from realistic and synthetic datasets. We also propose an inference strategy that can process extremely long videos through segment-wise estimation and seamless stitching. Comprehensive evaluations on multiple datasets reveal that DepthCrafter achieves state-of-the-art performance in open-world video depth estimation under zero-shot settings. Furthermore, DepthCrafter facilitates various downstream applications, including depth-based visual effects and conditional video generation.
Diffusion-based video generation technology has advanced significantly, catalyzing a proliferation of research in human animation. However, the majority of these studies are confined to same-modality driving settings, with cross-modality human body animation remaining relatively underexplored. In this paper, we introduce, an end-to-end audio-driven human animation framework that ensures hand integrity, identity consistency, and natural motion. The key design of CyberHost is the Region Codebook Attention mechanism, which improves the generation quality of facial and hand animations by integrating fine-grained local features with learned motion pattern priors. Furthermore, we have developed a suite of human-prior-guided training strategies, including body movement map, hand clarity score, pose-aligned reference feature, and local enhancement supervision, to improve synthesis results. To our knowledge, CyberHost is the first end-to-end audio-driven human diffusion model capable of facilitating zero-shot video generation within the scope of human body. Extensive experiments demonstrate that CyberHost surpasses previous works in both quantitative and qualitative aspects.
Human video generation task has gained significant attention with the advancement of deep generative models. Generating realistic videos with human movements is challenging in nature, due to the intricacies of human body topology and sensitivity to visual artifacts. The extensively studied 2D media generation methods take advantage of massive human media datasets, but struggle with 3D-aware control; whereas 3D avatar-based approaches, while offering more freedom in control, lack photorealism and cannot be harmonized seamlessly with background scene. We propose AMG, a method that combines the 2D photorealism and 3D controllability by conditioning video diffusion models on controlled rendering of 3D avatars. We additionally introduce a novel data processing pipeline that reconstructs and renders human avatar movements from dynamic camera videos. AMG is the first method that enables multi-person diffusion video generation with precise control over camera positions, human motions, and background style. We also demonstrate through extensive evaluation that it outperforms existing human video generation methods conditioned on pose sequences or driving videos in terms of realism and adaptability.
With the introduction of diffusion-based video generation techniques, audio-conditioned human video generation has recently achieved significant breakthroughs in both the naturalness of motion and the synthesis of portrait details. Due to the limited control of audio signals in driving human motion, existing methods often add auxiliary spatial signals to stabilize movements, which may compromise the naturalness and freedom of motion. In this paper, we propose an end-to-end audio-only conditioned video diffusion model named Loopy. Specifically, we designed an inter- and intra-clip temporal module and an audio-to-latents module, enabling the model to leverage long-term motion information from the data to learn natural motion patterns and improving audio-portrait movement correlation. This method removes the need for manually specified spatial motion templates used in existing methods to constrain motion during inference. Extensive experiments show that Loopy outperforms recent audio-driven portrait diffusion models, delivering more lifelike and high-quality results across various scenarios.
Variational Autoencoder (VAE), compressing videos into latent representations, is a crucial preceding component of Latent Video Diffusion Models (LVDMs). With the same reconstruction quality, the more sufficient the VAE's compression for videos is, the more efficient the LVDMs are. However, most LVDMs utilize 2D image VAE, whose compression for videos is only in the spatial dimension and often ignored in the temporal dimension. How to conduct temporal compression for videos in a VAE to obtain more concise latent representations while promising accurate reconstruction is seldom explored. To fill this gap, we propose an omni-dimension compression VAE, named OD-VAE, which can temporally and spatially compress videos. Although OD-VAE's more sufficient compression brings a great challenge to video reconstruction, it can still achieve high reconstructed accuracy by our fine design. To obtain a better trade-off between video reconstruction quality and compression speed, four variants of OD-VAE are introduced and analyzed. In addition, a novel tail initialization is designed to train OD-VAE more efficiently, and a novel inference strategy is proposed to enable OD-VAE to handle videos of arbitrary length with limited GPU memory. Comprehensive experiments on video reconstruction and LVDM-based video generation demonstrate the effectiveness and efficiency of our proposed methods.
The Diffusion Transformers Models (DiTs) have transitioned the network architecture from traditional UNets to transformers, demonstrating exceptional capabilities in image generation. Although DiTs have been widely applied to high-definition video generation tasks, their large parameter size hinders inference on edge devices. Vector quantization (VQ) can decompose model weight into a codebook and assignments, allowing extreme weight quantization and significantly reducing memory usage. In this paper, we propose VQ4DiT, a fast post-training vector quantization method for DiTs. We found that traditional VQ methods calibrate only the codebook without calibrating the assignments. This leads to weight sub-vectors being incorrectly assigned to the same assignment, providing inconsistent gradients to the codebook and resulting in a suboptimal result. To address this challenge, VQ4DiT calculates the candidate assignment set for each weight sub-vector based on Euclidean distance and reconstructs the sub-vector based on the weighted average. Then, using the zero-data and block-wise calibration method, the optimal assignment from the set is efficiently selected while calibrating the codebook. VQ4DiT quantizes a DiT XL/2 model on a single NVIDIA A100 GPU within 20 minutes to 5 hours depending on the different quantization settings. Experiments show that VQ4DiT establishes a new state-of-the-art in model size and performance trade-offs, quantizing weights to 2-bit precision while retaining acceptable image generation quality.
Creating editable videos that depict complex interactions between multiple objects in various artistic styles has long been a challenging task in filmmaking. Progress is often hampered by the scarcity of data sets that contain paired text descriptions and corresponding videos that showcase these interactions. This paper introduces a novel depth-conditioning approach that significantly advances this field by enabling the generation of coherent and diverse videos from just a single text-video pair using a pre-trained depth-aware Text-to-Image (T2I) model. Our method fine-tunes the pre-trained model to capture continuous motion by employing custom-designed spatial and temporal attention mechanisms. During inference, we use the DDIM inversion to provide structural guidance for video generation. This innovative technique allows for continuously controllable depth in videos, facilitating the generation of multiobject interactions while maintaining the concept generation and compositional strengths of the original T2I model across various artistic styles, such as photorealism, animation, and impressionism.
The advancement of autonomous driving technologies necessitates increasingly sophisticated methods for understanding and predicting real-world scenarios. Vision language models (VLMs) are emerging as revolutionary tools with significant potential to influence autonomous driving. In this paper, we propose the DriveGenVLM framework to generate driving videos and use VLMs to understand them. To achieve this, we employ a video generation framework grounded in denoising diffusion probabilistic models (DDPM) aimed at predicting real-world video sequences. We then explore the adequacy of our generated videos for use in VLMs by employing a pre-trained model known as Efficient In-context Learning on Egocentric Videos (EILEV). The diffusion model is trained with the Waymo open dataset and evaluated using the Fr\'echet Video Distance (FVD) score to ensure the quality and realism of the generated videos. Corresponding narrations are provided by EILEV for these generated videos, which may be beneficial in the autonomous driving domain. These narrations can enhance traffic scene understanding, aid in navigation, and improve planning capabilities. The integration of video generation with VLMs in the DriveGenVLM framework represents a significant step forward in leveraging advanced AI models to address complex challenges in autonomous driving.
Character animation is a transformative field in computer graphics and vision, enabling dynamic and realistic video animations from static images. Despite advancements, maintaining appearance consistency in animations remains a challenge. Our approach addresses this by introducing a training-free framework that ensures the generated video sequence preserves the reference image's subtleties, such as physique and proportions, through a dual alignment strategy. We decouple skeletal and motion priors from pose information, enabling precise control over animation generation. Our method also improves pixel-level alignment for conditional control from the reference character, enhancing the temporal consistency and visual cohesion of animations. Our method significantly enhances the quality of video generation without the need for large datasets or expensive computational resources.
Autonomous driving training requires a diverse range of datasets encompassing various traffic conditions, weather scenarios, and road types. Traditional data augmentation methods often struggle to generate datasets that represent rare occurrences. To address this challenge, we propose GenDDS, a novel approach for generating driving scenarios generation by leveraging the capabilities of Stable Diffusion XL (SDXL), an advanced latent diffusion model. Our methodology involves the use of descriptive prompts to guide the synthesis process, aimed at producing realistic and diverse driving scenarios. With the power of the latest computer vision techniques, such as ControlNet and Hotshot-XL, we have built a complete pipeline for video generation together with SDXL. We employ the KITTI dataset, which includes real-world driving videos, to train the model. Through a series of experiments, we demonstrate that our model can generate high-quality driving videos that closely replicate the complexity and variability of real-world driving scenarios. This research contributes to the development of sophisticated training data for autonomous driving systems and opens new avenues for creating virtual environments for simulation and validation purposes.
Video diffusion models are able to generate high-quality videos by learning strong spatial-temporal priors on large-scale datasets. In this paper, we aim to investigate whether such priors derived from a generative process are suitable for video recognition, and eventually joint optimization of generation and recognition. Building upon Stable Video Diffusion, we introduce GenRec, the first unified framework trained with a random-frame conditioning process so as to learn generalized spatial-temporal representations. The resulting framework can naturally supports generation and recognition, and more importantly is robust even when visual inputs contain limited information. Extensive experiments demonstrate the efficacy of GenRec for both recognition and generation. In particular, GenRec achieves competitive recognition performance, offering 75.8% and 87.2% accuracy on SSV2 and K400, respectively. GenRec also performs the best on class-conditioned image-to-video generation, achieving 46.5 and 49.3 FVD scores on SSV2 and EK-100 datasets. Furthermore, GenRec demonstrates extraordinary robustness in scenarios that only limited frames can be observed. Code will be available at https://github.com/wengzejia1/GenRec.
Fundus Fluorescein Angiography (FFA) is a critical tool for assessing retinal vascular dynamics and aiding in the diagnosis of eye diseases. However, its invasive nature and less accessibility compared to Color Fundus (CF) images pose significant challenges. Current CF to FFA translation methods are limited to static generation. In this work, we pioneer dynamic FFA video generation from static CF images. We introduce an autoregressive GAN for smooth, memory-saving frame-by-frame FFA synthesis. To enhance the focus on dynamic lesion changes in FFA regions, we design a knowledge mask based on clinical experience. Leveraging this mask, our approach integrates innovative knowledge mask-guided techniques, including knowledge-boosted attention, knowledge-aware discriminators, and mask-enhanced patchNCE loss, aimed at refining generation in critical areas and addressing the pixel misalignment challenge. Our method achieves the best FVD of 1503.21 and PSNR of 11.81 compared to other common video generation approaches. Human assessment by an ophthalmologist confirms its high generation quality. Notably, our knowledge mask surpasses supervised lesion segmentation masks, offering a promising non-invasive alternative to traditional FFA for research and clinical applications. The code is available at https://github.com/Michi-3000/Fundus2Video.