Distributed target blocks for fast autoregressive video generation

01 / Watch

Watch SpecAR finish first

Both clips start from the same prompt and opening frame. The race uses each method’s measured mean generation time, so SpecAR finishes while the target keeps generating. Select another scene and it plays automatically; switch to “Same frame” to compare the two outputs at matching video times.

02 / Mechanism

Where should the target blocks go?

Suppose we are generating a video of 36 blocks, but can afford to run the large model only seven times. Where should we spend those seven invocations?

We could concentrate them at the beginning, scatter them randomly, or decide as we go using a perceptual router. But these choices can leave long stretches where the smaller model generates without any intervention. And because each block becomes part of the context for the next, errors made early in such a stretch can propagate through everything that follows.

SpecAR takes a different approach: spread the target blocks as evenly as possible.

For 36 blocks and seven target invocations, this guarantees that the drafter never generates more than five consecutive blocks on its own. No other placement with the same budget can guarantee a shorter maximum run. The schedule is determined before generation begins, requires no training, and adds no per-block scoring overhead.

This is a guarantee about how often the target intervenes, not a guarantee of video quality. Whether such a simple schedule can actually outperform an adaptive router is an empirical question.

Try moving the target blocks below. Cluster them, scatter them, and see how the longest uninterrupted draft run changes.

Given a rollout of H blocks and a target budget k, SpecAR places target blocks at approximately equal intervals, including block 0. Exactly one model generates each block; its output is committed to a shared context for all later blocks. The worst draft run is at most ⌈H/k⌉ − 1 blocks, which is the theoretical minimum.

See the decoding diagram from the paper

03 / Results

Can a fixed schedule outperform an adaptive router?

Even spacing gives us a useful guarantee: the drafter can never run for too long without the target intervening. But minimizing uninterrupted draft runs does not necessarily produce better video. An adaptive router has access to information that a fixed schedule ignores. In principle, that should give it an advantage.

So we compared SpecAR against SDVG, which uses perceptual scores to decide when to invoke the target model.

The results were surprising. Across all three evaluated target budgets, SpecAR was both faster and better on the measured quality metrics. On a single H100, it ran 1.63–1.73× faster than SDVG on Krea Realtime 14B, and 1.10–1.84× faster on Helios 14B, while achieving higher VisionReward and VBench Quality-6 scores at every evaluated operating point.

For example, with seven target blocks on Krea Realtime, SpecAR finished in 42.9 seconds, compared with 70.3 seconds for SDVG. Its VisionReward score was also higher: 0.0972 versus 0.0886.

The router had more information and made more decisions. Neither translated into an advantage in these experiments.

04 / Quick start

Run SpecAR yourself

The code behind these results is publicly available. We've released SpecAR with support for Krea Realtime 14B and Helios 14B, allowing you to generate videos, adjust the target-block budget, and explore the speed–quality tradeoff on a single supported GPU.

You can run SpecAR directly from the command line or use the ComfyUI integration for a visual workflow. Both expose the same central idea: control how often the target model runs, without relying on an online router.

The released code and public weights run on one NVIDIA Hopper GPU (H100 80GB or H200 141GB) with Python 3.11. Choose a way to run it below. FlashAttention 3 is fetched prebuilt on first use.

Krea Realtime 14BHelios 14BComfyUI

Install

Copy

git clone https://github.com/statite-ai/SpecAR && cd SpecAR
python3.11 -m venv .venv-krea && source .venv-krea/bin/activate
pip install ".[krea]"
specar download --backbone krea

Generate

Copy

specar generate --backbone krea \
    --prompt "A papercraft coral reef full of colorful fish" \
    --k 2 --out reef.mp4

About 50 GB download. A 26.6 s video (426 frames) generates in under a minute on an H100, after about 3 minutes of model loading.

--k is the number of target blocks: 2 is the fastest setting in the paper, 7 the default. Set the length with --seconds or try a curated prompt with --example cat; all options are in the README.

References

1. Hassan, H. (2026). SpecAR: Distributed Target Blocks for Fast Autoregressive Video Generation. Preprint. H100 benchmarks and routing ablations.
2. Hu, Y., & Zhang, J. (2026). Speculative Decoding for Autoregressive Video Generation. arXiv:2604.17397. SDVG baseline.
3. Millon, E., & Krea AI. (2025). Krea Realtime 14B: Real-Time, Long-Form AI Video Generation. Model description.
4. Yuan, S., et al. (2026). Helios: Real Real-Time Long Video Generation Model. arXiv:2603.04379.
5. Huang, X., et al. (2025). Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion. arXiv:2506.08009. Krea drafter.
6. Huang, Z., et al. (2024). VBench: Comprehensive Benchmark Suite for Video Generative Models. CVPR 2024.
7. Xu, J., et al. (2026). VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation. AAAI 2026.

B200 measurements on this page average four Krea prompts and are separate from the preprint.

Cite this work

@misc{hassan2026specar,
  title  = {SpecAR: Distributed Target Blocks for Fast Autoregressive Video Generation},
  author = {Hassan, Hudeifa},
  year   = {2026},
  note   = {Preprint},
  url    = {https://statite.ai/specar}
}