The generation of temporally consistent, high-fidelity driving videos over extended horizons presents a fundamental challenge in autonomous driving world modeling. Existing approaches often suffer from error accumulation and feature misalignment due to inadequate decoupling of spatio-temporal dynamics and limited cross-frame feature propagation mechanisms. To address these limitations, we present STAGE (Streaming Temporal Attention Generative Engine), a novel auto-regressive framework that pioneers hierarchical feature coordination and multi-phase optimization for sustainable video synthesis. To achieve high-quality long-horizon driving video generation, we introduce Hierarchical Temporal Feature Transfer (HTFT) and a novel multi-stage training strategy. HTFT enhances temporal consistency between video frames throughout the video generation process by modeling the temporal and denoising process separately and transferring denoising features between frames. The multi-stage training strategy is to divide the training into three stages, through model decoupling and auto-regressive inference process simulation, thereby accelerating model convergence and reducing error accumulation. Experiments on the Nuscenes dataset show that STAGE has significantly surpassed existing methods in the long-horizon driving video generation task. In addition, we also explored STAGE's ability to generate unlimited-length driving videos. We generated 600 frames of high-quality driving videos on the Nuscenes dataset, which far exceeds the maximum length achievable by existing methods.

1. Bibliographic Information

• Title: STAGE: A Stream-Centric Generative World Model for Long-Horizon Driving-Scene Simulation
• Authors: Jiamin Wang, Yichen Yao, Xiang Feng, Hang Wu, Yaming Wang, Qingqiu Huang, Yuexin Ma, Xinge Zhu.
  • Affiliations: The authors are affiliated with various institutions, including what appears to be academic labs (indicated by superscript 1) and corporate research groups (indicated by superscript 2). The specific affiliations are not fully detailed in the provided text, but the authors' work is situated at the intersection of academic and industry research in autonomous driving.
• Journal/Conference: The paper is currently hosted on arXiv, a repository for electronic preprints of scientific papers. This indicates it has not yet completed a formal peer-review process for a specific conference or journal, or it is a pre-publication version.
• Publication Year: 2024 (based on the arXiv submission date).
• Abstract: The paper introduces STAGE (Streaming Temporal Attention Generative Engine), an auto-regressive framework designed to generate high-fidelity, temporally consistent driving videos over long durations. The authors identify error accumulation and feature misalignment in existing methods as key challenges. To solve this, STAGE proposes two main innovations: a Hierarchical Temporal Feature Transfer (HTFT) mechanism to improve temporal consistency by sharing features between frames during the denoising process, and a multi-stage training strategy to accelerate convergence and reduce error accumulation. Experiments on the NuScenes dataset show that STAGE significantly outperforms prior methods in long-horizon video generation and demonstrates the ability to generate videos of "unlimited-length," with a successful generation of a 600-frame sequence.
• Original Source Link:

  • arXiv Page: https://arxiv.org/abs/2506.13138v3

  • PDF Link: https://arxiv.org/pdf/2506.13138v3.pdf

  • Publication Status: Preprint.

    ---

2. Executive Summary

• Background & Motivation (Why):

  • Core Problem: Generating realistic, long, and controllable driving videos is a fundamental task for training and validating autonomous driving systems. Such simulations can replace expensive and time-consuming real-world data collection. However, existing methods struggle.
  • Existing Gaps:
    1. One-shot models (which generate the whole video at once) are computationally expensive and cannot flexibly adjust the video length.
    2. Auto-regressive models (which generate the video chunk-by-chunk or frame-by-frame) suffer from error accumulation, where small imperfections in one generated frame are amplified in subsequent frames, leading to a rapid degradation of video quality and consistency.
    3. Many models lack effective mechanisms for propagating features across long time spans, resulting in poor temporal consistency (e.g., flickering, objects appearing/disappearing unnaturally).
  • Fresh Angle: STAGE introduces a stream-centric approach, generating videos frame-by-frame auto-regressively. This offers maximum flexibility in video length. To combat the typical downsides of this approach, it introduces two novel techniques: HTFT to enforce temporal consistency at a deep feature level, and a multi-stage training strategy to make the model robust to its own generative errors.

• Main Contributions / Findings (What):

  1. STAGE Framework: A novel streaming-based generative world model that can produce high-quality driving videos of arbitrary length by generating one frame at a time.

  2. Hierarchical Temporal Feature Transfer (HTFT): A new mechanism that enhances temporal consistency. It transfers denoising features from previous frames to the current frame at each step of the generation process, providing strong temporal guidance.

  3. Multi-Stage Training Strategy: A progressive three-stage training scheme that decouples model components for faster convergence and simulates the auto-regressive inference process during training. This directly tackles the train-inference discrepancy, reducing error accumulation.

  4. State-of-the-Art Performance: The paper demonstrates that STAGE significantly outperforms existing methods on long-horizon video generation tasks on the NuScenes dataset.

  5. Unlimited-Length Generation: The authors showcase the model's ability to generate a 600-frame high-quality video, far longer than the training data sequences, by integrating a predictive model for future control signals.

     ---

3. Prerequisite Knowledge & Related Work

• Foundational Concepts:

  • World Model: A type of AI model that learns an internal representation of an environment. It can be used to simulate future states of that environment based on past observations and current actions. In this context, the "world" is a driving scene, and the model learns to predict how the scene will evolve over time, generating future video frames.
  • Auto-regressive Model: A model that generates a sequence of data one step at a time, where each new step is conditioned on the previous steps. For video, this means generating frame $T$ based on frames $1, 2, ..., T-1$. While flexible, this process can cause errors to compound.
  • Diffusion Models: A class of powerful generative models that have become state-of-the-art for image and video synthesis. They work by first adding random noise to a real image in a series of steps and then training a neural network (typically a U-Net) to reverse this process, i.e., to denoise the image and reconstruct it. To generate a new image, the model starts with pure noise and progressively denoises it. STAGE is built upon the architecture of Stable Diffusion, a well-known latent diffusion model.
  • Temporal Consistency: In video generation, this is the quality of objects and scenes appearing to move and change smoothly and logically over time. A lack of consistency results in flickering, objects popping in and out of existence, or illogical motion.
  • Latent Space: A compressed, lower-dimensional representation of data. Instead of generating high-resolution pixels directly, many models (like Stable Diffusion) first encode an image into a smaller latent representation, perform the generation process in this efficient latent space, and then decode the final latent back into a full-resolution image.

• Previous Works:

  • The paper situates itself against two main classes of video generation methods for driving:
    1. One-Shot Generation: Models like MagicDriveDiT use a Transformer-based architecture (DiT) to generate an entire long video in a single forward pass.
       • Limitation: These are resource-intensive (high VRAM usage) and inflexible—they are trained to produce a fixed-length video and cannot easily generate longer or shorter sequences.
    2. Auto-regressive Generation: Models like Vista, Drive-WM, and GenAD generate long videos through multiple inference steps, typically creating a short video chunk conditioned on the previous one.
       • Limitation: They suffer from error accumulation. The quality of each new chunk depends on the previously generated one, and any imperfections are carried forward and amplified, leading to significant quality degradation over long horizons.
  • Other related works like StreamingT2V and FIFO-diffusion have explored streaming generation for general-purpose videos but are not specifically designed for the complex, controllable driving scenes that require precise layout control (e.g., via HD maps and bounding boxes).

• Differentiation:

  • STAGE is a streaming auto-regressive model, generating frame-by-frame, which offers more flexibility than chunk-based auto-regressive models and one-shot models.

  • Unlike previous methods that struggle with temporal consistency, STAGE introduces HTFT, a novel feature-sharing mechanism that operates during the denoising process, explicitly linking the generation of consecutive frames at a deep level.

  • To combat error accumulation, STAGE uses a multi-stage training approach, specifically Stage 3, which forces the model to learn from its own imperfect outputs, a key difference from standard training that always uses perfect ground-truth data.

    ---

4. Methodology (Core Technology & Implementation)

The core of STAGE is a streaming auto-regressive framework built on a diffusion model. It generates a video frame by frame, $I_0, I_1, ..., I_T$. To generate frame $I_T$, it uses information from previous frames and control signals.

As shown in Image 2, the overall architecture generates frames in a streaming manner. The key innovations, HTFT and the multi-stage training, are central to its operation. The diagram also illustrates how infinite generation is achieved by looping the output back as an input condition.

• A. Problem Formulation and Overview:

  • Goal: Generate a long, high-quality video sequence $\{I_0, I_1, ..., I_T\}$ given an initial frame $I_0$ and corresponding control signals like HD Map $H_0$ and Bounding Box layout $B_0$.
  • Process: The model operates in latent space. At each step $T$, it generates the latent feature for frame $I_T$ conditioned on:
    1. The latent feature of the previous frame, $I_{T-1}$ (the condition frame).
    2. The latent feature of the first frame, $I_0$ (the anchor frame), to maintain global consistency.
    3. Control signals (predicted HD map and bounding boxes for frame $T$).
  • The StreamingBuffer is a memory component that stores features from past frames to be used by HTFT.

• B. Hierarchical Temporal Feature Transfer (HTFT):

  • Principle: The key insight is that during the denoising process of a diffusion model, the intermediate features within its U-Net architecture contain rich spatial and contextual information. HTFT leverages this by transferring these features across time to guide the generation of the current frame, ensuring it is consistent with past frames. The "hierarchical" nature comes from transferring features across both time (from previous frames) and denoising steps (at the same level of noise).
  • Steps & Procedures: During the generation of frame $T$ at denoising step $t$, the model performs the following:
    1. Feature Storage: A StreamingBuffer, implemented as a FIFO (First-In, First-Out) queue of length $N=10$, stores the intermediate U-Net features from the last 10 generated frames. For each frame, it stores features from every denoising step.
    2. Feature Selection: Instead of using features from all 10 previous frames (which would be computationally expensive), a frame-skipping strategy is used. It selects features from a predefined set of past frames, specifically the 1st, 5th, and 10th previous frames ($S = \{-1, -5, -10\}$).
    3. Feature Fusion: The selected historical features are fused with the current frame's features using cross-attention.
  • Mathematical Formulas: Let $f_t^T$ be the intermediate feature for the current frame $T$ at denoising step $t$.
    1. The StreamingBuffer stores features from previous frames: $$F_s = \mathrm{FIFO}(\{f_t^{T-1}, f_t^{T-2}, \dots, f_t^{T-N}\})$$ Here, $F_s$ is the buffer containing features from the same denoising step $t$ for the past $N$ frames.
    2. Select features from frames at indices specified by $S$: $F_e = \bigcup_{x \in S} F_s[x]$
    3. Process the current frame's feature $f_t^T$ to create a query for attention: $$g_t^T = \mathrm{Linear}(\mathrm{GroupNorm}(f_t^T))$$
    4. Use cross-attention to inject temporal information from past frames ($F_e$) into the current frame's context: $$H_t^T = \mathrm{CrossAttn}(g_t^T, F_e)$$
    5. Fuse the temporal context back into the main feature stream via a residual connection: $$F_{fused} = \mathrm{Dropout}(f_t^T) + \mathrm{Linear}(H_t^T)$$ This $F_{fused}$ then replaces $f_t^T$ in the U-Net, guiding the denoising process with strong temporal priors.

• C. Training Strategy: A three-stage strategy is proposed to progressively build the model's capabilities and address the train-inference mismatch.

  • Stage 1: Streaming Learning: The model is trained for basic frame-by-frame generation using ground-truth conditional inputs. The HTFT mechanism is disabled. This stage establishes a solid foundation for video generation.
  • Stage 2: HTFT Learning: The weights of the base model from Stage 1 are frozen. Only the newly added network components for HTFT (e.g., the cross-attention and linear layers) are trained. This allows the model to efficiently learn how to transfer temporal features without disturbing the already learned generation capabilities.
  • Stage 3: Learning Through Simulating Inference: This is the crucial stage for reducing error accumulation. The model is trained on sequences where the condition frame (the previous frame $I_{T-1}$) is not the ground-truth image but the model's own generated output from the previous step. This forces the model to become robust to the kinds of artifacts and quality degradation it will encounter during actual long-horizon inference. To keep this computationally feasible, inference is run once per sequence, not for every single frame during training.
  • Data Augmentation: To improve robustness, several augmentations are used. The condition frame is degraded using Discrete Cosine Transform (DCT) filtering to simulate generation artifacts. Random dropout and noise are also added to the conditioning inputs.
  • Loss Function: A weighted L2 distance between the model's prediction of the original image and the actual original image. $$L = \mathbb { E } _ { x _ { t } , x _ { 0 } , t } \left[ \| W _ { a u x } \odot ( x _ { 0 } - \epsilon _ { \theta } ( x _ { t } , t ) ) \| ^ { 2 } \right]$$
    • Symbol Explanation:
      • $x_0$: The original, clean image latent.
      • $x_t$: The noised image latent at step $t$.
      • $\epsilon_{\theta}(x_t, t)$: The model's prediction of $x_0$ given $x_t$. The paper's notation is slightly ambiguous; typically $\epsilon_\theta$ predicts the noise, not $x_0$, but the goal is equivalent.
      • $W_{aux}$: A per-pixel weight map.
      • $\odot$: Element-wise multiplication. The key is the auxiliary weight map $W_{aux}$, which is designed to prioritize important objects. $$w_{ij}' = \begin{cases} k / p_{ij}^c & (i, j) \in \text{foreground polygon} \\ 1 / (H * W)^c & (i, j) \in \text{background polygon} \end{cases}$$
    • The weight w'_{ij} for a pixel (i,j) is inversely proportional to the size $p_{ij}$ of the foreground object it belongs to, raised to the power $c$. This gives higher weights to smaller objects, ensuring they are generated with high fidelity. The weights are then normalized. Object shapes are defined by convex hulls of their 3D bounding boxes for more accurate area calculation.

• D. Infinite Generation: To generate videos longer than the provided ground-truth data, STAGE needs a way to predict future control signals (HD maps, bounding boxes). The paper integrates the MILE model for this purpose. At each step $T$ beyond the available data, MILE predicts the ego vehicle's future state and the positions of other agents, which are then converted into the HD map and bounding box conditions for STAGE to generate the next frame, $I_{T+1}$. This creates a closed-loop simulation engine.

---

5. Experimental Setup

• Datasets: The NuScenes dataset, a large-scale dataset for autonomous driving research. It contains 700 training and 150 validation sequences, each about 20 seconds long. The authors upsample the annotations to 12 Hz to create a denser video sequence.

• Evaluation Metrics:

  1. FID (Fréchet Inception Distance): Measures the quality and realism of individual generated images. It calculates the "distance" between the feature distributions of real and generated images. Features are extracted from a pre-trained Inception-v3 network. A lower FID means the generated images are statistically more similar to real images.
     • Conceptual Definition: FID assesses how "real" the generated images look by comparing their statistical properties (mean and covariance) in a deep feature space to those of real images.
     • Mathematical Formula: $$\mathrm{FID}(x, g) = \| \mu_x - \mu_g \|_2^2 + \mathrm{Tr}(\Sigma_x + \Sigma_g - 2(\Sigma_x \Sigma_g)^{1/2})$$
     • Symbol Explanation:
       • x, g: The sets of real and generated images, respectively.
       • $\mu_x, \mu_g$: The mean of the Inception-v3 feature vectors for real and generated images.
       • $\Sigma_x, \Sigma_g$: The covariance matrices of the feature vectors for real and generated images.
       • $\mathrm{Tr}(\cdot)$: The trace of a matrix (sum of diagonal elements).
  2. FVD (Fréchet Video Distance): The video-domain equivalent of FID. It measures both the per-frame image quality and the temporal consistency of a video. It uses features from a pre-trained video recognition model (the paper uses VideoGPT) to compare distributions. A lower FVD indicates better video quality and motion dynamics.
     • Conceptual Definition: FVD evaluates video realism by comparing the statistical properties of generated videos to real videos in a feature space that captures both appearance and motion.
     • Mathematical Formula: The formula is identical to FID, but the features are extracted from a video model instead of an image model.
     • Symbol Explanation: Same as FID, but $\mu$ and $\Sigma$ now represent the mean and covariance of feature distributions of video clips.

• Baselines: The paper compares STAGE with several state-of-the-art models:

  • DriveDreamer, Drive-WM, MagicDrive, DreamForge, Vista: These are primarily auto-regressive or world-model-based approaches.

  • MagicDriveDiT: A leading one-shot generation model.

    ---

6. Results & Analysis

• Core Results: The main quantitative results are presented in Table I, comparing STAGE to baselines on short-horizon (16 frames) and long-horizon (up to 240 frames) generation.

  (Manual transcription of Table I) TABLE I: COMPARISON WITH STATE-OF-THE-ART METHODS IN SHORT-HORIZON AND LONG-HORIZON GENERATION. LOWER IS BETTER FOR ALL METRICS.

  Setting	Methods	FID↓	FVD↓
  short term	DriveDreamer [14]	52.60	452.00
  	Drive-WM [6]	15.80	122.70
  	MagicDrive [4]	16.20	217.90
  	MagicDriveDiT [3]	20.91	94.84
  	DreamForge [18]	16.00	224.80
  	Ours	11.04	242.79
  long term	MagicDriveDiT [3]	-	585.89
  	Vista [5]	90.55	626.58
  	Ours	23.70	280.34

  • Long-Horizon Analysis: STAGE achieves state-of-the-art results, dramatically outperforming all competitors. Its FVD of 280.34 is less than half that of MagicDriveDiT (585.89), and its FID of 23.70 is far superior to Vista's (90.55). This demonstrates the effectiveness of the proposed methods (HTFT and multi-stage training) in preventing error accumulation and maintaining temporal consistency over long sequences.
  • Short-Horizon Analysis: STAGE achieves the best FID (11.04), indicating superior single-frame image quality. However, its FVD (242.79) is higher than some specialized short-video models like MagicDriveDiT (94.84). The authors attribute this to competitors being based on Stable Video Diffusion, a model pre-trained specifically for short video generation. Crucially, STAGE's FVD degrades only slightly from short-horizon (242.79) to long-horizon (280.34), while competitors see a massive drop in performance, highlighting STAGE's exceptional temporal stability.

• Qualitative Analysis:

  • Image 3: In a long-horizon comparison with Vista, STAGE's generated frames remain sharp and coherent up to frame 201, while Vista's output becomes blurry and distorted due to error accumulation. This visually confirms STAGE's stability.

    该图像是一个插图，展示了基于Nuscenes数据集的长时序驾驶视频生成比较，分别对比了Vista方法、本文提出的STAGE方法和真实视频（GT）在不同时间帧（T=40至T=200）下的生成效果，凸显了STAGE在细节和场景一致性上的优势。

  • Image 4: In a short-horizon comparison, STAGE demonstrates better condition-following. While both models produce high-quality images, the ego car in Vista's video moves forward when it should be stationary, whereas STAGE correctly follows the ground truth motion.

    该图像是一个插图，展示了Vista方法、本文方法（Our）和真实地面真值（GT）在不同时间步（T=1,4,8,12,15）的道路场景视频帧对比，突出本文方法在长时序视频生成中保持场景一致性的优越性。

  • Image 6: The model accurately generates vehicles within the specified bounding box locations, both during the day and at night, proving its strong controllability.

    该图像是论文中用于展示边界框控制效果的插图。左侧为条件输入边界框，中央是模型生成的结果，右侧为真实场景图像，展示了模型在不同场景下的生成能力。

• Ablation Study: The ablation study in Table II validates the contribution of each training stage.

  (Manual transcription of Table II) TABLE II: THE ABLATION STUDIES AT DIFFERENT TRAINING STAGES. LOWER IS BETTER FOR ALL METRICS.

  	FID↓	FVD↓
  Stage 1	17.09	508.29
  Stage 2	11.90	245.11
  Stage 3	11.04	242.79

  • Stage 1 -> Stage 2: Introducing HTFT (Stage 2) causes a massive improvement. FVD drops from 508.29 to 245.11, and FID improves from 17.09 to 11.90. This proves that HTFT is critical for achieving both high image quality and temporal consistency.
  • Stage 2 -> Stage 3: Simulating inference during training (Stage 3) provides a further, noticeable improvement in both FID and FVD. This confirms that making the model robust to its own errors is effective in reducing the train-inference gap and enhancing final performance.

• Generation Exceeding Original Data Length:

  • Image 5: This figure demonstrates the model's most ambitious claim: near-infinite generation. It shows frames from a continuous 600-frame sequence (equivalent to 50 seconds at 12Hz). The image quality and scene coherence remain remarkably high even at frame 600, far beyond the ~20-second length of the training sequences. This is made possible by the Stage 4 model that uses MILE to predict future conditions.

    该图像是两组驾驶场景视频帧的对比示意图，展示了在不同时间步（T=120至T=600）下模型生成的长时序驾驶视频效果，突出STAGE方法在长时段内保持场景一致性和细节的能力。

---

7. Conclusion & Reflections

• Conclusion Summary: The paper successfully introduces STAGE, a streaming generative world model that sets a new state-of-the-art in long-horizon driving video generation. Its frame-by-frame approach provides great flexibility, while the novel contributions of Hierarchical Temporal Feature Transfer (HTFT) and a multi-stage training strategy effectively solve the long-standing problems of temporal inconsistency and error accumulation in auto-regressive models. The demonstrated ability to generate a 600-frame high-quality video highlights its potential for creating extensive, realistic simulation data for autonomous driving.

• Limitations & Future Work:

  • The paper does not explicitly state limitations. However, a potential weakness is the reliance on an external model (MILE) for infinite generation. The quality of the generated video in this closed-loop setting is ultimately capped by the predictive accuracy of MILE. If the predicted conditions become unrealistic, the generated video will likely follow suit.
  • Future work is centered on exploring the full potential of "infinite length" video generation, which the authors have already demonstrated as a proof-of-concept. Refining the closed-loop simulation to handle more complex and long-term interactions would be a natural next step.

• Personal Insights & Critique:

  • Strengths: The design of STAGE is both elegant and pragmatic. The HTFT mechanism is a clever way to enforce temporal consistency at a fundamental level of the generation process. The multi-stage training strategy, particularly Stage 3, is a practical and effective solution to the train-inference mismatch that plagues many generative models. The results are highly compelling and clearly demonstrate a significant leap forward for long-video generation in this domain.
  • Critical Reflections:
    • Scalability and Efficiency: While more flexible than one-shot models, the frame-by-frame generation with cross-attention in HTFT could still be computationally intensive for real-time or very large-scale simulation. The paper does not provide detailed metrics on generation speed.
    • Generalization to "Out-of-Distribution" Scenarios: The model is trained on NuScenes data. Its ability to generate truly novel or rare "edge-case" scenarios (which is a key promise of simulation) is not fully explored. The closed-loop system with MILE might tend to generate scenarios that are statistically similar to the training data.
    • Physical Realism: While the videos look visually coherent, a rigorous analysis of the underlying physical plausibility (e.g., vehicle dynamics, collision avoidance) is not presented. For autonomous driving simulation, this is a critical aspect that goes beyond visual appeal.