The frontier of video generation is sliding from "it moves" to "it is sharp": 720p is becoming the entry ticket, and 2K is the next battleground. But every step up in resolution inflates the quadratic cost of full attention in the Transformer; the subtler damage is that attention spreads over more and more redundant tokens, diluting the learning signal that actually matters. Prism, a framework submitted to arXiv on October 4 by Tencent Hunyuan together with Fudan University and Zhejiang University (arXiv:2610.05416), upgrades sparse attention from a common inference-acceleration trick into a native structure for training, aimed squarely at the 2K training bill of joint video-audio generation.
Joint video-audio generation, attention that treats everything alike
Joint video-audio generation means the model produces picture and sound in a single denoising pass. The paper starts from a structural fact that is easy to miss: in joint video-audio data, cross-modal interactions are inherently concentrated around sound-producing regions — that is where audio actually exerts influence. Full attention treats all of this alike, spreading compute across the entire spatiotemporal volume; once resolution climbs to 2K and token counts explode, redundant tokens dilute the learning signal and disrupt pretrained priors. The authors also note that existing sparse-attention methods mostly target training-free inference acceleration, and none of them are designed around this region-based structure of joint video-audio data.
Let the attention block shape follow the content
Prism's core idea is to partition attention by content. It first organizes the token sequence into spatiotemporal macro-zones, so the attention structure can adapt to local content. For each zone it estimates two signals: the variance of video features along the channel dimension, and the feature norms from audio-to-video cross-attention — the former captures where visual content changes sharply, the latter measures how strongly audio influences each visual region. From these signals, Prism dynamically assigns a tailored block shape to every zone, applying finer partitioning along axes of rapid visual variation and strong audio-visual coupling. Tokens inside each block stay semantically coherent, so block-level features capture both visual content and joint video-audio interaction patterns. A hybrid block-selection strategy then decides per-query sparsity.
2.5x faster training, and quality goes up
The headline result: versus full attention, Prism achieves a 2.5x training speedup while surpassing it in generation quality. The code and preview weights are open-sourced under MIT (GitHub: Tencent-Hunyuan/Prism), with 300+ stars at the time of writing. The preview model supports native joint video-audio generation and training at 720p, 1080p, and 2K, with structured audio control via <music>, <sfx>, and <speech> tags; the alpha checkpoint targets stability, the beta targets motion, and a stronger Prism-pro is on the roadmap.
Hardware bar, and what to watch
Cold water first: the repository is blunt about hardware. Native 720p training needs at least 32 NVIDIA 80 GB GPUs, and 1080p/2K each need at least 64; for inference, 720p runs on a single 80 GB card while 1080p and 2K require 4 or more. This is a framework for research teams with clusters, not a consumer toy.
Two things are worth noting. First, training-side sparsity and inference-side sparsity are complementary tracks: the industry has mostly pursued training-free sparse acceleration at inference time to save deployment cost; Prism bakes sparsity into training itself to save training budget — the two can stack. Second, joint audio-video generation is becoming an open race, and the MIT license plus the full pipeline (data preprocessing, bucketing, FSDP training, inference scripts) continues Tencent Hunyuan's open-source cadence in video generation.
So the next leg of the resolution race may hinge not on bigger clusters but on smarter attention structures. Once the attention bill for 2K training can be cut by more than half, "natively high-resolution" stops being a mega-lab privilege and becomes a replicable engineering problem.