Sol Video Inference Engine

Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation

A training-free acceleration framework that composes cache, sparse attention, token pruning, quantization, and kernel fusion into deployment-ready stacks for video diffusion inference.

NVIDIA SANA Team*Codex*GPT-Image2Human authors: Yitong Li, Junsong Chen, Haopeng Li, Haozhe Liu, Jincheng Yu, Ligeng Zhu, Ping Luo, Song Han, Enze Xie.NVIDIA Research, Efficient AI Team & Singapore Lab.

View pricing PAPER CODE MINIMAX H3 H3 SUPER ACCELERATION SOL-H3 LTX-2.5

LTX-2.3 PROGRESS DEMO

Try the experience

Generate with Sol Engine

Sol Engine Demo

Describe what you want to create0 / 2000

Reference model

Acceleration mode

Key Features

Instance Specificity

Model × hardware × configuration aware acceleration

The framework optimizes each concrete deployment instance rather than assuming a single transferable acceleration recipe.

Full-Stack Techniques

Cache, sparse attention, token pruning, quantization, and kernel fusion

Sol Video Inference Engine organizes five broadly applicable acceleration techniques into a deployment-ready full-stack framework.

Local-to-Global Workflow

Parallel skill agents with global integration

Per-technique local tuning provides strong starting points that are later composed through a global integrator and validation loop.

Efficiency Gains

2×–3× end-to-end speedup

The full framework substantially reduces latency across Cosmos3-Super, LTX-2.3, and SANA-Video while preserving perceptual quality.

Abstract

Sol Video Inference Engine is a training-free, agent-native acceleration framework for video diffusion inference that composes cache, sparse attention, token pruning, quantization, and kernel fusion into deployment-ready stacks tailored to each model, hardware target, and serving configuration. Across Cosmos3-Super, LTX-2.3, and SANA-Video, the framework delivers more than 2× end-to-end speedup while maintaining near-lossless visual quality.

Efficiency at a Glance

Sol Video Inference Engine exposes each technique as a bounded design space. Parallel skill agents test technique applicability, an agent integrator composes selected candidates into a deployment stack, and a human validator provides efficiency-quality feedback.

Sol Video Inference Engine acceleration figure

Demos

Cosmos3-Super

SGLANG · 1× baselineSOL-ENGINE · 2.27× accelerated

LTX-2.3

SGLANG · 1× baselineSOL-ENGINE · 2.38× accelerated

SANA-Video

SGLANG · 1× baselineSOL-ENGINE · 2.77× accelerated

BibTeX

Cite the Sol Video Inference Engine

@misc{li2026solvideoinferenceengine,
  title={Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation},
  author={Yitong Li and Junsong Chen and Haopeng Li and Haozhe Liu and Jincheng Yu and Ligeng Zhu and Ping Luo and Song Han and Enze Xie},
  year={2026},
  eprint={2606.23743},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.23743}
}