Helios: Real Real-Time Long Video Generation Model

Summary

14B autoregressive diffusion with Unified History Injection (native T2V/I2V/V2V), Easy Anti-Drifting (relative RoPE, first-frame anchor, frame-aware corrupt), Deep Compression Flow (multi-term memory patchification + pyramid UniPC) and adversarial hierarchical distillation to 3 steps. 19.5 FPS on one H100; HeliosBench 240 prompts at 81/240/720/1440 frames. Distilled scores 6.00 short / 6.94 long, beating prior real-time 1.3B methods while matching same-size base quality. Code https://github.com/PKU-YuanGroup/Helios; project https://pku-yuangroup.github.io/Helios-Page/.

Keywords

helios · video-generation · diffusion · real-time · anti-drifting · wan · bytedance · pku

Topics

video generation, autoregressive diffusion, real-time inference

Research notes

• Primary: arxiv abs (cs.CV). CC BY 4.0 on HTML. PKU / ByteDance China / Canva / Chengdu Anu Intelligence. Correspondence yuanshenghai@stu.pku.edu.cn; Li Yuan project lead. Code https://github.com/PKU-YuanGroup/Helios (2,051 stars at check); project page repo Helios-Page (23 stars). HF paper page 190 upvotes; 8 linked models not copied into hf_* fields. Discord posted HF papers URL. HeliosBench is 240 prompts, not a hosted corpus, so no datasets_local row. License field left blank per catalog convention.