Awesome World Model Hub Papers · Datasets · Projects
← Back to papers

Real Real-Time Long Video Generation Model

arXiv 26.3 2026 43.6 method

TLDR

Helios is a 14B autoregressive diffusion model for real-time long video generation, achieving 19.5 FPS on a single H100 without standard acceleration or anti-drifting heuristics.

Reasoning

The paper presents strong empirical results in real-time long video generation with novel training strategies for drifting and efficiency optimizations. However, it does not frame the work as a world model or address world dynamics prediction, limiting relevance to the specified keywords.

Read-first score

Read-first score 43.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 0.

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 25%
81

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Methodology quality 25%
60

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,experiment

Topical relevance 42%
0

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 517.

Keyword Scores

world model
0
world simulator
0
generative world model
0
interactive world model
0
video world model
0
world dynamics prediction
0
model-based reinforcement learning world model
0

Deep Analysis

Innovations

  • Robustness to long-video drifting without commonly used anti-drifting heuristics such as self-forcing, error-banks, or keyframe sampling
  • Real-time generation without standard acceleration techniques such as KV-cache, sparse/linear attention, or quantization
  • Training without parallelism or sharding frameworks, enabling image-diffusion-scale batch sizes while fitting up to four 14B models within 80 GB of GPU memory
  • Unified input representation that natively supports T2V, I2V, and V2V tasks
  • Training strategies that explicitly simulate drifting during training and eliminate repetitive motion at its source
  • Heavy compression of historical and noisy context and reduction of sampling steps for efficiency
  • Infrastructure-level optimizations that accelerate both inference and training while reducing memory consumption

Methodology

Helios is a 14B autoregressive diffusion model with a unified input representation supporting text-to-video, image-to-video, and video-to-video tasks. It uses training strategies that simulate drifting during training to mitigate long-video drifting, and compresses historical and noisy context while reducing sampling steps for efficiency. Infrastructure optimizations accelerate both inference and training while reducing memory consumption.

Key Results

Helios achieves 19.5 FPS on a single NVIDIA H100 GPU for minute-scale video generation, matching the quality of a strong baseline and consistently outperforming prior methods on both short- and long-video generation.

Tags