Speculative Decoding for Autoregressive Video Generation
Summary
SDVG adapts speculative decoding to autoregressive video diffusion, using an image-quality router to achieve up to 2.09× speed-up with 95.7% quality retention on MovieGenVideoBench.
View Cached Full Text
Cached at: 04/22/26, 06:17 AM
Paper page - Speculative Decoding for Autoregressive Video Generation
Source: https://huggingface.co/papers/2604.17397
Abstract
Speculative decoding is adapted to autoregressive video diffusion through a quality-based routing mechanism that maintains high visual quality while achieving significant speedup.
Autoregressive video diffusionis emerging as a promising paradigm for streaming video synthesis, withstep distillationserving as the primary means of accelerating inference. Whetherspeculative decoding, the dominant acceleration strategy for large language models, can be effectively adapted to autoregressive video generation remains an open question, because video blocks are continuous spatiotemporal tensors with no token-level distribution for exact rejection sampling. We introduce SDVG, which bringsspeculative decodingto block-basedautoregressive video diffusionby replacing token verification with an image-quality router. A 1.3B drafter proposes candidate blocks via fourdenoising steps; each block is VAE-decoded and scored byImageRewardusingworst-frame aggregation--taking the minimum per-frame reward to catch single-frame artifacts that averaging would mask. Blocks scoring above a fixed threshold tau are accepted into the 14B target’sKV cache; the rest are regenerated by the target. Two additional design choices prove critical: the first block is always force-rejected to anchor scene composition, and tau serves as a single knob that traces a smooth quality-speedPareto frontier. On 1003 MovieGenVideoBench prompts (832x480), SDVG retains 98.1% of target-onlyVisionRewardquality (0.0773 vs. 0.0788) at a 1.59x speedup with tau=-0.7, and reaches 2.09x at 95.7% quality retention--while consistently outperforming draft-only generation by over +17%. The framework is training-free, requires no architectural changes, and can be seamlessly integrated into existing autoregressive video generation pipelines.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2604\.17397
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2604.17397 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2604.17397 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2604.17397 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators
OPSD-V improves few-step autoregressive video diffusion models by using real long-video data as temporal context during training, providing dense trajectory-level supervision that enhances visual quality and motion dynamics without altering inference mechanisms.
@rohanpaul_ai: Spatially Speculative Decoding (SSD) sped up autoregressive image models up to 13.28X by predicting image rows in paral…
Spatially Speculative Decoding (SSD) accelerates autoregressive image models by predicting entire rows in parallel using small helper networks, achieving up to 13.28x speedup while maintaining benchmark performance.
AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding
AngelSpec introduces a unified training and inference framework for speculative decoding that jointly optimizes autoregressive multi-token prediction and block-parallel diffusion drafters to handle heterogeneous real-world workloads. Experiments on the Hy3 model series show up to 2.4x speedup over autoregressive decoding and 11.8% higher throughput than DFlash.
Long Video Generation (4 minute read)
The article introduces A²RD, a novel architecture for generating consistent long videos using agentic autoregressive diffusion. It proposes a Retrieve–Synthesize–Refine–Update cycle and a new benchmark, LVBench-C, to address semantic drift in long-horizon video synthesis.
What is Speculative Decoding? (trending on paperswithco.de) [R]
Speculative decoding is an inference optimization technique that uses a fast draft model to propose future tokens verified in parallel by a larger model, improving LLM generation speed. The article highlights its trending status on Papers with Code and a recent SGLang blog post about state-of-the-art latencies using DFlash models.