HiStream: Efficient High-Resolution Video Generation via Redundancy-Eliminated Streaming

TL;DR

HiStream tackles the prohibitive compute cost of high-resolution video diffusion by eliminating redundancy across spatial, temporal, and timestep dimensions. It introduces a dual-resolution caching scheme, an anchor-guided sliding window, and an optional asymmetric denoising pathway to deliver state-of-the-art 1080p video quality with substantial speedups (up to 107.5x over the baseline). The method stacks progressive upscaling, fixed-attention caches, and a fixed-anchor attention strategy to maintain temporal coherence without growing memory usage. Thorough experiments on 1080p benchmarks show strong qualitative fidelity and robust efficiency gains, with HiStream+ offering an even more aggressive speed-accuracy trade-off. The work highlights remaining bottlenecks in VAE decoding and training memory, outlining concrete directions for future enhancement of practical high-resolution video diffusion systems.

Abstract

High-resolution video generation, while crucial for digital media and film, is computationally bottlenecked by the quadratic complexity of diffusion models, making practical inference infeasible. To address this, we introduce HiStream, an efficient autoregressive framework that systematically reduces redundancy across three axes: i) Spatial Compression: denoising at low resolution before refining at high resolution with cached features; ii) Temporal Compression: a chunk-by-chunk strategy with a fixed-size anchor cache, ensuring stable inference speed; and iii) Timestep Compression: applying fewer denoising steps to subsequent, cache-conditioned chunks. On 1080p benchmarks, our primary HiStream model (i+ii) achieves state-of-the-art visual quality while demonstrating up to 76.2x faster denoising compared to the Wan2.1 baseline and negligible quality loss. Our faster variant, HiStream+, applies all three optimizations (i+ii+iii), achieving a 107.5x acceleration over the baseline, offering a compelling trade-off between speed and quality, thereby making high-resolution video generation both practical and scalable.

Figures (9)

• Figure 1: HiStream delivers high-resolution (1080p) autoregressive video generation, offering up to $76.2\times$ faster denoising than the bi-directional baseline and $2.5\times$ faster than self-forcing, all while preserving the highest preference ratings. Furthermore, our faster variant, HiStream+, achieves a remarkable $107.5\times$ acceleration against the baseline and $3.5\times$ faster than self-forcing.
• Figure 2: Pipeline details. Illustration of the synergy among our three core efficiency mechanisms. (Left) The Anchor-Guided Sliding Window strategy ensures robust temporal scalability by generating the video in fixed-size chunks. Each generation step maintains a constant attention context by combining tokens from the persistent content anchor (the first frame, dark green) and recent historical context (local frames, dark blue), thereby avoiding the growth of the KV cache over time. (Right) The Dual-Resolution Caching accelerates synthesis by adopting a two-stage process: initial low-resolution denoising is followed by high-resolution refinement. Crucially, the final high-resolution output updates $\text{KV}_{\text{high}}$ and is subsequently downsampled to update $\text{KV}_{\text{low}}$, ensuring spatial consistency for subsequent chunks. With Asymmetric Denoising, subsequent chunks use half as many denoising steps as the first chunk.
• Figure 3: Results visualization of different timesteps. (Left) A notable quality gap exists at Step 1 between frame 9 (chunk 1) and frame 17 (chunk 2). The fully denoised cache from chunk 1 provides high-value context, enabling chunk 2 to achieve a near-final quality (comparable to Step 3) in just one step. (Right) Blur Trial: We also test robustness by blurring the outputs of Steps 1 and 2 (via down/upsampling) in the 4-step process. The final result remains nearly indistinguishable from the Baseline Setting.
• Figure 4: Visualization of temporal attention. (Right) Attention during autoregressive generation naturally focuses on the initial frame and the two most recent neighboring frames. (Left) This visualization demonstrates that dropping intermediate, less important frames has a negligible impact, even without retraining. Conversely, dropping the initial frame causes catastrophic quality degradation.
• Figure 5: Qualitative comparisons. We visualize high-resolution videos generated by HiStream (Ours) against those by Wan2.1 wan2025, Self Forcing huang2025self, LTX HaCohen2024LTXVideo, and FlashVideo zhang2025flashvideo. The videos generated by HiStream exhibit the highest visual fidelity and the cleanest texture, free from spurious patterns or visible artifacts. Best viewed ZOOMED-IN.