Controllable generative models for images and videos have seen significant success, yet 3D scene generation, especially in unbounded scenarios like autonomous driving, remains underdeveloped. Existing methods lack flexible controllability and often rely on dense view data collection in controlled environments, limiting their generalizability across common datasets (e.g., nuScenes). In this paper, we introduce MagicDrive3D, a novel framework for controllable 3D street scene generation that combines video-based view synthesis with 3D representation (3DGS) generation. It supports multi-condition control, including road maps, 3D objects, and text descriptions. Unlike previous approaches that require 3D representation before training, MagicDrive3D first trains a multi-view video generation model to synthesize diverse street views. This method utilizes routinely collected autonomous driving data, reducing data acquisition challenges and enriching 3D scene generation. In the 3DGS generation step, we introduce Fault-Tolerant Gaussian Splatting to address minor errors and use monocular depth for better initialization, alongside appearance modeling to manage exposure discrepancies across viewpoints. Experiments show that MagicDrive3D generates diverse, high-quality 3D driving scenes, supports any-view rendering, and enhances downstream tasks like BEV segmentation, demonstrating its potential for autonomous driving simulation and beyond.

Instruction tuning of Large Vision-language Models (LVLMs) has revolutionized the development of versatile models with zero-shot generalization across a wide range of downstream vision-language tasks. However, the diversity of different training tasks from various sources and formats would lead to inevitable task conflicts, where different tasks conflict for the same set of model parameters, resulting in sub-optimal instruction-following abilities. To address that, we propose the Mixture of Cluster-conditional LoRA Experts (MoCLE), a novel Mixture of Experts (MoE) architecture designed to activate task-customized model parameters based on instruction clusters. A separate universal expert is further incorporated to improve generalization abilities of MoCLE for novel instructions. Extensive experiments on InstructBLIP and LLaVA demonstrate the effectiveness of MoCLE.

Zhili Liu, Yunhao Gou, Kai Chen, Lanqing Hong, Jiahui Gao, Fei Mi, Yu Zhang, Zhenguo Li, Xin Jiang, Qun Liu, James Kwok. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Yunhao Gou, Hansi Yang, Zhili Liu, Kai Chen, Yihan Zeng, Lanqing Hong, Zhenguo Li, Qun Liu, Bo Han, James Kwok, Yu Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large Vision-Language Models (LVLMs) have received widespread attentions for advancing the interpretable self-driving. Existing evaluations of LVLMs primarily focus on multi-faceted capabilities in natural circumstances, lacking automated and quantifiable assessment for self-driving, let alone the severe road corner cases. In this work, we propose CODA-LM, the very first benchmark for the automatic evaluation of LVLMs for self-driving corner cases. We adopt a hierarchical data structure and prompt powerful LVLMs to analyze complex driving scenes and generate high-quality pre-annotations for the human annotators, while for LVLM evaluation, we show that using the text-only large language models (LLMs) as judges reveals even better alignment with human preferences than the LVLM judges. Moreover, with our CODA-LM, we build CODA-VLM, a new driving LVLM surpassing all open-sourced counterparts on CODA-LM. Our CODA-VLM performs comparably with GPT-4V, even surpassing GPT-4V by +21.42% on the regional perception task. We hope CODA-LM can become the catalyst to promote interpretable self-driving empowered by LVLMs.

Despite remarkable achievements in video synthesis, achieving granular control over complex dynamics, such as nuanced movement among multiple interacting objects, still presents a significant hurdle for dynamic world modeling, compounded by the necessity to manage appearance and disappearance, drastic scale changes, and ensure consistency for instances across frames. These challenges hinder the development of video generation that can faithfully mimic real-world complexity, limiting utility for applications requiring high-level realism and controllability, including advanced scene simulation and training of perception systems. To address that, we propose TrackDiffusion, a novel video generation framework affording fine-grained trajectory-conditioned motion control via diffusion models, which facilitates the precise manipulation of the object trajectories and interactions, overcoming the prevalent limitation of scale and continuity disruptions. A pivotal component of TrackDiffusion is the instance enhancer, which explicitly ensures inter-frame consistency of multiple objects, a critical factor overlooked in the current literature. More-over, we demonstrate that generated video sequences by our TrackDiffusion can be used as training data for visual per-ception models. To the best of our knowledge, this is the first work to apply video diffusion models with tracklet conditions and demonstrate that generated frames can be beneficial for improving the performance of object trackers. 1

Current perceptive models heavily depend on resource-intensive datasets, prompting the need for innovative solutions. Leveraging recent advances in diffusion models, synthetic data, by constructing image inputs from various annotations, proves beneficial for downstream tasks. While prior methods have separately addressed generative and perceptive models, DetDiffusion, for the first time, harmonizes both, tackling the challenges in generating effective data for perceptive models. To enhance image generation with perceptive models, we introduce perception-aware loss (P.A. loss) through segmentation, improving both quality and controllability. To boost the performance of specific perceptive models, our method customizes data augmentation by extracting and utilizing perception-aware attribute (P.A. Attr) during generation. Experimental results from the object detection task highlight DetDiffusion's superior performance, establishing a new state-of-the-art in layout-guided generation. Furthermore, image syntheses from DetDiffusion can effectively augment training data, significantly enhancing downstream detection performance.

Recent advancements in diffusion models have significantly enhanced the data synthesis with 2D control. Yet, precise 3D control in street view generation, crucial for 3D perception tasks, remains elusive. Specifically, utilizing Bird's-Eye View (BEV) as the primary condition often leads to challenges in geometry control (e.g., height), affecting the representation of object shapes, occlusion patterns, and road surface elevations, all of which are essential to perception data synthesis, especially for 3D object detection tasks. In this paper, we introduce MagicDrive, a novel street view generation framework, offering diverse 3D geometry controls including camera poses, road maps, and 3D bounding boxes, together with textual descriptions, achieved through tailored encoding strategies. Besides, our design incorporates a cross-view attention module, ensuring consistency across multiple camera views. With MagicDrive, we achieve high-fidelity street-view image & video synthesis that captures nuanced 3D geometry and various scene descriptions, enhancing tasks like BEV segmentation and 3D object detection. Project Website: https://flymin.github.io/magicdrive

The rapid development of large language models (LLMs) has not only provided numerous opportunities but also presented significant challenges. This becomes particularly evident when LLMs inadvertently generate harmful or toxic content, either unintentionally or because of intentional inducement. Existing alignment methods usually direct LLMs toward the favorable outcomes by utilizing human-annotated, flawless instruction-response pairs. Conversely, this study proposes a novel alignment technique based on mistake analysis, which deliberately exposes LLMs to erroneous content to learn the reasons for mistakes and how to avoid them. In this case, mistakes are repurposed into valuable data for alignment, effectively helping to avoid the production of erroneous responses. Without external models or human annotations, our method leverages a model's intrinsic ability to discern undesirable mistakes and improves the safety of its generated responses. Experimental results reveal that our method outperforms existing alignment approaches in enhancing model safety while maintaining the overall utility.

Diffusion models have attracted significant attention due to the remarkable ability to create content and generate data for tasks like image classification. However, the usage of diffusion models to generate the high-quality object detection data remains an underexplored area, where not only image-level perceptual quality but also geometric conditions such as bounding boxes and camera views are essential. Previous studies have utilized either copy-paste synthesis or layout-to-image (L2I) generation with specifically designed modules to encode the semantic layouts. In this paper, we propose the GeoDiffusion, a simple framework that can flexibly translate various geometric conditions into text prompts and empower pre-trained text-to-image (T2I) diffusion models for high-quality detection data generation. Unlike previous L2I methods, our GeoDiffusion is able to encode not only the bounding boxes but also extra geometric conditions such as camera views in self-driving scenes. Extensive experiments demonstrate GeoDiffusion outperforms previous L2I methods while maintaining 4x training time faster. To the best of our knowledge, this is the first work to adopt diffusion models for layout-to-image generation with geometric conditions and demonstrate that L2I-generated images can be beneficial for improving the performance of object detectors.

Masked Autoencoder (MAE) has demonstrated superior performance on various vision tasks via randomly masking image patches and reconstruction. However, effective data augmentation strategies for MAE still remain open questions, different from those in contrastive learning that serve as the most important part. This paper studies the prevailing mixing augmentation for MAE. We first demonstrate that naïve mixing will in contrast degenerate model performance due to the increase of mutual information (MI). To address, we propose homologous recognition, an auxiliary pretext task, not only to alleviate the MI increasement by explicitly requiring each patch to recognize homologous patches, but also to perform object-aware self-supervised pre-training for better downstream dense perception performance. With extensive experiments, we demonstrate that our proposed Mixed Autoencoder (MixedAE) achieves the state-of-the-art transfer results among masked image modeling (MIM) augmentations on different downstream tasks with significant efficiency. Specifically, our MixedAE outperforms MAE by +0.3% accuracy, +1.7 mIoU and +0.9 AP on ImageNet-1K, ADE20K and COCO respectively with a standard ViT-Base. Moreover, MixedAE surpasses iBOT, a strong MIM method combined with instance discrimination, while accelerating training by 2×. To our best knowledge, this is the very first work to consider mixing for MIM from the perspective of pretext task design. Code will be made available.

Recent years have witnessed a few attempts of vision transformers for single image super-resolution (SISR). Since the high resolution of intermediate features in SISR models increases memory and computational requirements, efficient SISR transformers are more favored. Based on some popular transformer backbone, many methods have explored reasonable schemes to reduce the computational complexity of the self-attention module while achieving impressive performance. However, these methods only focus on the performance on the training platform (e.g., Pytorch/Tensorflow) without further optimization for the deployment platform (e.g., TensorRT). Therefore, they inevitably contain some redundant operators, posing challenges for subsequent deployment in real-world applications. In this paper, we propose a deployment-friendly transformer unit, namely UFONE (i.e., UnFolding ONce is Enough), to alleviate these problems. In each UFONE, we introduce an Inner-patch Transformer Layer (ITL) to efficiently reconstruct the local structural information from patches and a Spatial-Aware Layer (SAL) to exploit the long-range dependencies between patches. Based on UFONE, we propose a Deployment-friendly Inner-patch Transformer Network (DITN) for the SISR task, which can achieve favorable performance with low latency and memory usage on both training and deployment platforms. Furthermore, to further boost the deployment efficiency of the proposed DITN on TensorRT, we also provide an efficient substitution for layer normalization and propose a fusion optimization strategy for specific operators. Extensive experiments show that our models can achieve competitive results in terms of qualitative and quantitative performance with high deployment efficiency.

Self-supervised learning (SSL), especially contrastive methods, has raised attraction recently as it learns effective transferable representations without semantic annotations. A common practice for self-supervised pre-training is to use as much data as possible. For a specific downstream task, however, involving irrelevant data in pre-training may degenerate the downstream performance, observed from our extensive experiments. On the other hand, for existing SSL methods, it is burdensome and infeasible to use different downstream-task-customized datasets in pre-training for different tasks. To address this issue, we propose a novel SSL paradigm called Scalable Dynamic Routing (SDR), which can be trained once and deployed efficiently to different downstream tasks with task-customized pre-trained models. Specifically, we construct the SDRnet with various sub-nets and train each sub-net with only one subset of the data by data-aware progressive training. When a downstream task arrives, we route among all the pre-trained sub-nets to get the best along with its corresponding weights. Experiment results show that our SDR can train 256 sub-nets on ImageNet simultaneously, which provides better transfer performance than a unified model trained on the full ImageNet, achieving state-of-the-art (SOTA) averaged accuracy over 11 downstream classification tasks and AP on PASCAL VOC detection task.

Face image super resolution (face hallucination) usu-ally relies on facial priors to restore realistic details and preserve identity information. Recent advances can achieve impressive results with the help of GAN prior. They ei-ther design complicated modules to modify the fixed GAN prior or adopt complex training strategies to finetune the generator. In this work, we propose a generative and controllable face SR framework, called GCFSR, which can re-construct images with faithful identity information without any additional priors. Generally, GCFSR has an encoder-generator architecture. Two modules called style modu-lation and feature modulation are designed for the multi-factor SR task. The style modulation aims to generate real-istic face details and the feature modulation dynamically fuses the multi-level encoded features and the generated ones conditioned on the upscaling factor. The simple and elegant architecture can be trained from scratch in an end-to-end manner. For small upscaling factors (≤8), GCFSR can produce surprisingly good results with only adversar-ialloss. After adding L1 and perceptual losses, GCFSR can outperform state-of-the-art methods for large upscalingfac-tors (16, 32, 64). During the test phase, we can modulate the generative strength via feature modulation by changing the conditional upscaling factor continuously to achieve various generative effects. Code is available at https://github.com/hejingwenhejingwen/GCFSR

Autonomous driving has attracted much attention over the years but turns out to be harder than expected, probably due to the difficulty of labeled data collection for model training. Self-supervised learning (SSL), which leverages unlabeled data only for representation learning, might be a promising way to improve model performance. Existing SSL methods, however, usually rely on the single-centric-object guarantee, which may not be applicable for multi-instance datasets such as street scenes. To alleviate this limitation, we raise two issues to solve: (1) how to define positive samples for cross-view consistency and (2) how to measure similarity in multi-instance circumstances. We first adopt an IoU threshold during random cropping to transfer global-inconsistency to local-consistency. Then, we propose two feature alignment methods to enable 2D feature maps for multi-instance similarity measurement. Addition-ally, we adopt intra-image clustering with self-attention for further mining intra-image similarity and translation-invariance. Experiments show that, when pre-trained on Waymo dataset, our method called Multi-instance Siamese Network (MultiSiam) remarkably improves generalization ability and achieves state-of-the-art transfer performance on autonomous driving benchmarks, including Cityscapes and BDD100K, while existing SSL counterparts like MoCo, MoCo-v2, and BYOL show significant performance drop. By pre-training on SODA10M, a large-scale autonomous driving dataset, MultiSiam exceeds the ImageNet pre-trained MoCo-v2, demonstrating the potential of domain-specific pre-training. Code will be available at https://github.com/KaiChen1998/MultiSiam.

Learning a large-scale regression model has proven to be one of the most successful approaches for visual tracking as in recent correlation filter (CF)- based trackers. Different from the conventional CF-based algorithms in which the regression model is solved based on circulant training samples, we propose learning linear regression models via a single-convolutional layer with the gradient descent (GD) technique. In our convolution-based approach, the samples are cropped from an image in a sliding-window manner rather than being circularly shifted from one base sample. As a result, the abundant background context in the images can be fully exploited to learn a robust tracker. The proposed tracker is based on two independent regression models: a holistic regression model and a texture regression model. The holistic regression model is trained based on the entire object patch to predict the object location, whereas the texture regression model is trained based on the local object textures. The foreground map outputted by the texture regression model is not only helpful to boost the location prediction in the case of large variations, but is also an important clue for estimating the object size. With the foreground map outputted by the texture regression model, we are able to estimate the object size by optimizing a novel objective function based on object-background contrast. Our extensive experiments on four popular visual tracking datasets OTB-50, OTB-100, VOT-2016, and TempleColor have proved that the proposed algorithm achieves outstanding performance and outperforms most CF-based trackers.

The main challenges of visual object tracking arise from the arbitrary appearance of the objects that need to be tracked. Most existing algorithms try to solve this problem by training a new model to regenerate or classify each tracked object. As a result, the model needs to be initialized and retrained for each new object. In this paper, we propose to track different objects in an object-independent approach with a novel two-flow convolutional neural network (YCNN). The YCNN takes two inputs (one is an object image patch, the other is a larger searching image patch), then outputs a response map which predicts how likely and where the object would appear in the search patch. Unlike the object-specific approaches, the YCNN is actually trained to measure the similarity between the two image patches. Thus, this model will not be limited to any specific object. Furthermore, the network is end-to-end trained to extract both shallow and deep dedicated convolutional features for visual tracking. And once properly trained, the YCNN can be used to track all kinds of objects without further training and updating. As a result, our algorithm is able to run at a very high speed of 45 frames-per-second. The effectiveness of the proposed algorithm can also be proved by the experiments on two popular data sets: OTB-100 and VOT-2014.

Recently, discriminatively learned correlation filters (DCF) has attracted much attention in visual object tracking community. The success of DCF is potentially attributed to the fact that a large number of samples are utilized to train the ridge regression model and predict the location of an object. To solve the regression problem in an efficient way, these samples are all generated by circularly shifting from a searching patch. However, these synthetic samples also induce some negative effects which weaken the robustness of DCF based trackers. In this paper, we propose a new approach to learn the regression model for visual tracking with single convolutional layer. Instead of learning the linear regression model in a closed form, we try to solve the regression problem by optimizing a onechannel- output convolution layer with gradient descent (GD). In particular, the kernel size of the convolution layer is set to the size of the object. Contrary to DCF, it is possible to incorporate all "real" samples clipped from the whole image. A critical issue of the GD approach is that most of the convolutional samples are negative and the contribution of positive samples will be suppressed. To address this problem, we propose a novel objective function to eliminate easy negatives and enhance positives. We perform extensive experiments on four widely-used datasets: OTB-100, OTB-50, TempleColor, and VOT-2016. The results show that the proposed algorithm achieves outstanding performance and outperforms most of the existing DCF based algorithms.