Stable Video Infinity: Infinite-Length Video Generation (Oct 2025)

Описание: Title: Stable Video Infinity: Infinite-Length Video Generation with Error Recycling (Oct 2025)
Link: http://arxiv.org/abs/2510.09212v1
Date: October 2025

Summary:
The paper introduces Stable Video Infinity (SVI), a method for generating infinite-length videos with temporal consistency, plausible scene transitions, and controllable storylines. It addresses the error accumulation problem by using Error-Recycling Fine-Tuning, which recycles Diffusion Transformer self-generated errors into supervisory prompts, enabling the model to identify and correct its own errors through closed-loop recycling.

Key Topics:
Video Generation
Diffusion Transformer
Error Recycling
Long-form Video
Temporal Consistency
Autoregressive Models

Chapters:
00:00 - Introduction to SVI
00:11 - The Error Accumulation Problem
00:30 - The SVI Solution: ERFT
00:53 - The Training Test Hypothesis Gap
01:30 - Air Recycling Fine-Tuning (ERFT)
02:13 - Diffusion Transformers
03:09 - Predictive vs. Conditional Error
03:46 - Cross-Clip Conditional Error
04:48 - The Vicious Cycle of Errors
05:28 - Goal of ERFT
06:14 - Error Replay Memory Banks
07:16 - Bidirectional Error Curation
08:15 - Ablation Studies
08:48 - Selective Sampling Strategy
09:56 - SVI Results: Consistency, Creativity, Control
10:33 - Solving Long-Term Drift
11:18 - SVI Film: Creative Scene Transitions
12:13 - Versatility: Audio and Motion Capture
13:17 - ERFT for Large Language Models (LLMs)
14:17 - Future: Real-time Video Generation