Controllable streaming video generation offers a promising route toward interactive game simulation, yet existing methods remain far from practical gameplay. They often fail to jointly achieve high visual fidelity, high frame rates, and low interaction latency, resulting in poor responsiveness and degraded user experience. We present Streaming-AAA, a new paradigm for controllable infinite generation of AAA-quality game videos. Our system generates 1080p gameplay videos at 43 FPS with 198 ms end-to-end latency interaction on merely NVIDIA A800 GPUs, while supporting long-horizon play without severe temporal degradation or collapse. Streaming-AAA is built upon three key designs. First, we redesign the VAE architecture for efficient high-resolution decoding, enabling real-time synthesis at 1080p. Second, we pretrain the diffusion model using diffusion forcing and significantly extend the temporal training horizon, allowing the model to capture long-term dependencies and maintain persistent memory. Third, we perform post-training distillation and memory compression to improve inference efficiency while retaining visual quality and temporal consistency. Experiments show that Streaming-AAA substantially improves visual fidelity, frame rate, latency, and long-term stability over prior controllable video generation approaches, enabling a significantly more playable experience for interactive AAA game video generation.
Figures & tables
Figure 1: Overview of Streaming-AAA: Streaming AAA game video generation framework applied to Genshin Impact , demonstrating high-resolution, high-frame-rate, and low-latency gameplay synthesis. Speed and latency are tested on NVIDIA A800 GPUs.
Figure 2: Overview of the ADCS. Designed to capture rich features of AAA games—including diverse landscapes, dynamic interactions (e.g., climbing, swimming), and photorealistic lighting—the system achieves a daily throughput of >10M frames and densely covers KM 2 s of virtual terrain within several days.
Dataset Overview
Total Videos
∼22 K
Video Duration
1 min
Total Frames
∼40 M
Resolution
1920×1080
Frame Rate
30
Coverage Area
> 5 KM 2
Table 1: Statistics of the Collected AAA Game Dataset. Our dataset provides high-resolution video paired with rich metadata.
Figure 3: Overview of the inference pipeline. During inference, the 3D UNet-based Diffusion Model takes noise as input, injects control signals via cross-attention, and propagates historical frame information through temporal attention. After multiple denoising steps, the reconstructed latents are decoded by the VAE decoder to produce the generated video frames.
Figure 4: Overview of the Hybrid KV-Cache Memory. First row shows the Vanilla KV-cache array, second row shows the proposed Hybrid KV-Cache array.
Figure 5: Parallelized Teacher-Forcing Inference. We merge the first denoise step and kv cache writing process to accelerate the overall inference procedure.
Figure 6: Qualitative results of our method on Genshin Impact . The top eleven rows show rollouts from in-domain frames, while the bottom two rows show rollouts from out-of-domain frames.
Figure 7: Qualitative comparison with state-of-the-art approaches on Genshin Impact . We show our method and the competing approaches on four different cases over a one-minute rollout; rows are methods and columns are timestamps.
Figure 8: The results of playing continuously for one hour using our model.
Generation Quality
Efficiency
Method
Subject Consist.
Background Consist.
Motion Smooth.
Imaging Quality
Average
Resolution
Frame Rate ( ↑ )
Latency ( ↓ )
Lingbot
0.689
0.876
0.983
0.438
0.747
720p
2.6 FPS (8GPUs)
8200 ms
YUME
0.716
0.860
0.985
0.639
0.800
720p
6.7 FPS
6600 ms
Matrix Game 3.0
0.690
0.887
0.987
0.666
0.807
720p
6.1 FPS
9800 ms
HY-World1.5
0.755
0.906
0.988
0.589
0.810
480p
2.1 FPS
11300 ms
Ours
0.761
0.910
0.982
0.677
0.832
1080p
43.0 FPS
198 ms
Table 2: Quantitative and efficiency comparison with state-of-the-art methods. We compare our method against existing state-of-the-art approaches on several VBench metrics, together with output resolution, frame rate, and action-to-pixel latency. Best result per column is in bold.
Figure 9: Ablation study on Genshin Impact . We compare the pre-trained model against post-training with a short and an extended training horizon across three scenarios: Normal navigation from in-domain (Case 1) and out-of-domain (Case 2) starting frames; Static standstill for 20 s and 60 s; and Memory retention after turning the camera away for 4 s/8 s and back.
Recent advancements in generative models are pushing video generation toward unbounded streaming audio-video generation for real-time interactive worlds. However, existing benchmarks primarily evaluate completed sequences and struggle to capture streaming properties. To bridge this gap, we introduce StreamAV-Bench, the first comprehensive benchmark tailored for streaming audio-video generation. StreamAV-Bench establishes a unified evaluation framework, including the progressive track for instruction adherence and long-horizon stability, and the interactive track for interactive response and state retention and reuse. With expert-verified evaluation cases across 32 fine-grained dimensions, we conduct an extensive evaluation of 13 representative systems. Our analysis reveals that current models suffer from temporal drift in progressive generation and responsiveness bottlenecks during interactive control. Based on a comprehensive failure analysis, we share insights to advance the development of native joint audio-video streaming models.
Autoregressive video diffusion models generate frames based on new observations and inputs, making them a promising medium for interactive visual experiences. However, existing models are orders of magnitude too slow for fluid interaction. This paper introduces FastFrames, a family of fast autoregressive video diffusion models that achieves interactive latency without compromising quality. We profile the inference pipeline of state-of-the-art models and find that operations often treated as fixed implementation details, such as VAE decoding and KV cache updates, in fact account for over half of end-to-end inference latency. We challenge the assumption that these components are indispensable and replace them with lightweight alternatives. Our 1.3B variant runs at 36 ms per frame, or 28 responses per second (RPS), on a single B200 GPU while matching the generation quality of a state-of-the-art model that is nearly 8x slower. Our 14B variant runs at 45 ms per frame, or 22 RPS, on an 8x B200 GPU node and yields significantly better visual quality and realism. We show that FastFrames enables real-time interaction in applications such as camera-conditioned and physics-conditioned video generation.
Recent advances in interactive video generations have demonstrated diffusion model's potential as world models by capturing complex physical dynamics and interactive behaviors. However, existing interactive world models depend on bidirectional attention and lengthy inference steps, severely limiting real-time performance. Consequently, they are hard to simulate real-world dynamics, where outcomes must update instantaneously based on historical context and current actions. To address this, we present Matrix-Game 2.0, an interactive world model generates long videos on-the-fly via few-step auto-regressive diffusion. Our framework consists of three key components: (1) A scalable data production pipeline for Unreal Engine and GTA5 environments to effectively produce massive amounts (about 1200 hours) of video data with diverse interaction annotations; (2) An action injection module that enables frame-level mouse and keyboard inputs as interactive conditions; (3) A few-step distillation based on the casual architecture for real-time and streaming video generation. Matrix Game 2.0 can generate high-quality minute-level videos across diverse scenes at an ultra-fast speed of 25 FPS. We open-source our model weights and codebase to advance research in interactive world modeling.