Tencent's MovieRoll streams a minute of AI video on 78% less cache
A Tencent Hunyuan team says its MovieRoll framework holds a 60-second single-shot generation together while using up to 78% less KV-cache memory than full-history caching.

Key takeaways
- Tencent Hunyuan's MovieRoll generates up to 60 seconds of video in one streaming pass, built on HunyuanVideo-1.5-480P-T2V.
- Its HieraKV compression cuts KV-cache memory by up to 78%, from 48.7 to 10.8 GiB at 18 cached chunks on one NVIDIA H20 GPU, and peak denoising-step latency with guidance from 33.8 to 15.8 seconds.
- On HeliosBench the 5-step MovieRoll-Distilled model posts the highest total scores at 10, 30 and 60 seconds, and a 59-professional user study preferred it over every baseline, with single-event win rates of 67.0% to 90.3%.
- On eight NVIDIA H20 GPUs a 60-second clip still takes 152.2 seconds to generate, or 9.47 FPS, against the 24 FPS the authors call real time, and the base model is too unstable at 60 seconds to be tested there.
Tencent Hunyuan's MovieRoll keeps a single generated shot running for a full minute, and it does it by letting the model forget most of what it no longer needs. The framework, posted to arXiv on 16 November 2026, is built on Tencent's own HunyuanVideo-1.5 model and trades a smaller memory footprint for the long horizons where streaming video generators usually drift, stall or run out of room.
The paper is a research report from a group spanning Tencent Hunyuan, Peking University and the Hong Kong University of Science and Technology, and every number in it is the authors' own, measured on their own hardware. No one outside that group has reproduced these results yet, so read the claims below as a team's account of its own work rather than a settled benchmark.
What is MovieRoll?
MovieRoll is what its authors call a base-to-distilled framework: one slow, high-quality model and one fast, few-step model trained from it. The slow one, MovieRoll-Base, is a multi-step generator trained with a method the paper calls Window-wise Teacher Forcing (WinTF), which denoises a rolling window of video chunks together instead of one at a time. The fast one, MovieRoll-Distilled, is squeezed down from the base with a training recipe called CuLT-DMD until it generates video in five denoising steps with no guidance.
Both are built on HunyuanVideo-1.5-480P-T2V, Tencent's text-to-video model, and both output 480p video at 24 frames per second. The framework generates from a single prompt or from an ordered list of prompts, so one video can carry several events, with a new prompt every ten seconds in the paper's headline example.
The problem MovieRoll is aimed at is specific. A streaming video model generates a chunk, then feeds it back as context for the next one, so the history it must attend to grows with every second. Training data rarely stretches past ten to fifteen seconds, and past that horizon the picture tends to wash out or wander. MovieRoll trains on clips of up to 15 seconds, then leans on its inference tricks to keep going to 60.
How the cache stops growing
The memory savings come from a scheme the authors call HieraKV, short for hierarchical KV-cache compression. A transformer keeps a key-value cache of everything it has already processed, and for long video that cache is most of the cost. HieraKV splits the model's layers into two groups: the shallow layers keep the full history, while the deeper layers keep only the sink chunk plus the most recent historical one.
Measured on MovieRoll-Base with guidance on a single NVIDIA H20 GPU, that split cuts KV-cache memory by up to 78%. At 18 cached chunks, the paper reports, the default allocation of six global and 48 local layers pulls memory down from 48.7 to 10.8 GiB and peak denoising-step latency from 33.8 to 15.8 seconds compared with caching everything in every layer. The authors say the compression improves generation quality rather than costing it, though that quality claim rests on their own scores.
Two infrastructure pieces support the same goal. FBCA, a block-wise causal attention kernel, had the lowest forward and training times in the paper's single-GPU comparison. PS-VAE, which splits video decoding across GPUs while keeping the decoder's state across chunks, made 1441-frame decoding possible at all: full-sequence decoding runs out of memory, and eight H20 GPUs gave the parallel decoder a 3.97 times speedup over one.
What the benchmarks and the user study show
The evaluation runs on HeliosBench, the benchmark from the Helios long-video project. The 5-step MovieRoll-Distilled model posts the highest total score of the compared methods at 10, 30 and 60 seconds for single-event generation and at 30 and 60 seconds for multi-event generation, against a field that includes CausVid, Self Forcing, Causal Forcing, Rolling Forcing, MemFlow, LongLive, Echo-Infinity, Krea and the two Helios variants.
A user study backs that up. Fifty-nine professionals compared videos made from the same prompt, with at least three experts judging each pair and 2,616 preference judgments collected in all. MovieRoll-Distilled won a majority against every baseline it was put against. In single-event comparisons the paper reports win rates from 67.0% to 90.3%, and in multi-event comparisons from 62.3% to 81.9%, counting ties in the totals. Those charts live on the project page rather than in the text, and they are the authors' own study, not a third-party evaluation.
A minute of video is still not real time
The catch is the speed. Even with SageAttention, a faster attention kernel, and eight NVIDIA H20 GPUs, generating a 60-second clip takes 152.2 seconds, which the paper converts to 9.47 frames per second. Real-time playback needs 24 frames per second, so the fastest configuration MovieRoll demonstrates is still about two and a half times too slow. On a single H20 GPU the distilled model manages 1.52 FPS, rising to 2.04 with SageAttention.
There is a second limit the authors state plainly: MovieRoll-Base is not stable enough at 60 seconds to be evaluated there, so only the distilled model is tested at the longest horizon. At 30 seconds the base model stays competitive, but a full minute is where the multi-step version gives out.
This is the same problem the whole long-video field is chewing on, and it is a different piece of it than the one SLOP TV covered on 15 November, when Geiru's context suite extended MiniMax H3 generation chains in ComfyUI. That work stretched an existing model's chains through a wrapper; MovieRoll changes the model's own caching and training so a single rollout survives the trip.
What it means for creators
You cannot run this on a laptop today. The code and the weights are out, but the README asks for FlashAttention-3, which needs a Hopper-class GPU such as an H100, and a full pipeline that fits on a single 96 GB card. CPU offload, Gradio support and consumer-GPU support (RTX 4090 and 5090) are all on the project's to-do list, not in it.
What the release does offer is a clear direction for anyone shipping long-form generation. The memory cost of a growing cache is the wall most streaming tools hit first, and MovieRoll's answer is to keep less of the past in more of the model. If that idea holds up outside Tencent's own hardware, it would lower the memory ceiling that keeps most long-form tools short. For now it is a research result, and the makers list improving inference efficiency among the work still to come.
MovieRoll's code and weights are on GitHub and Hugging Face under the Tencent Hunyuan Community License, with SGLang and vLLM-Omni backends supported from day one.
Sources
- arxiv.org - paper abstract, title, authors, submission date
- arxiv.org - full text: HieraKV 78%, latency 33.8 to 15.8s, speed 9.47 FPS, user study, limitations
- github.com - base model, model zoo, resolutions, license, hardware needs
- huggingface.co - the makers' own paper listing, dated 17 Nov 2026
- huggingface.co - where the weights are hosted
- tencent-hunyuan.github.io - project page with the user-study charts and speed tables
- aiweekly.co - a third-party write-up that restates the paper
- tldr.tech - a third-party listing of the GitHub release