Abstract
Generating long, multi-event videos with high temporal consistency remains a significant challenge for text-to-video (T2V) diffusion models. Current methods often suffer from "content drift," where character identity or environmental style shifts over time. We present SynCoS (Synchronized Coupled Sampling), a tuning-free inference framework that synchronizes denoising paths across extended sequences. By coupling local smoothness with global consistency through a three-stage iterative loop, SynCoS enables standard T2V models to generate 20+ second videos with complex, prompt-driven transitions while maintaining high fidelity to the original subject and style.
Introduction
Standard T2V models are typically constrained to short clips (2-5 seconds) due to the high computational cost of training on long-form data. Extending these models to longer durations often leads to two primary failures: content/style drift and the consistency-quality trade-off. In the former, characters morph into different individuals as the video progresses; in the latter, attempts to enforce global consistency often result in "collapsed" videos that lack dynamic motion. SynCoS addresses these issues by identifying and correcting misaligned denoising trajectories caused by varying prompts across different video segments.
Methodology: Synchronized Coupled Sampling
SynCoS introduces a three-stage iterative loop executed at every denoising step to ensure both local fluidity and global coherence.
The Three-Stage Loop
- Local Co-denoising: The sequence is divided into overlapping chunks. Each chunk is denoised using a specific Local Prompt (e.g., "he tilts his head"), and results are fused to ensure smooth frame-to-frame transitions.
- Global Refinement: The fused output is treated as a refinement source. Using Collaborative Score Distillation (CSD), the system synchronizes information across both adjacent and distant chunks to maintain a unified identity.
- Reversion: The refined output is reverted to the previous timestep, preparing the model for the next iteration.
Synchronization Anchors
To prevent artifacts and video collapse, SynCoS utilizes two key anchors:
- Grounded Timestep: The optimization stage is anchored to the same timestep as the sampling stage, ensuring both processes operate on the same level of generative detail.
- Fixed Baseline Noise: A single, fixed noise pattern is used across chunks during refinement to prevent the video from becoming a static image.
Experimental Results
SynCoS was evaluated on a benchmark of 48 long-video scenarios across multiple T2V backbones, including CogVideoX-2B and Open-Sora Plan. The framework demonstrated significant improvements over existing baselines like Gen-L-Video and FIFO-Diffusion.
- Temporal Consistency: On CogVideoX-2B, SynCoS achieved 88.92% subject consistency, a notable improvement over Gen-L-Video's 81.34%.
- Dynamic Motion: The model maintained a high "Dynamic Degree" of 85.42%, indicating that the generated videos remained active and followed complex action prompts without sacrificing stability.
- Qualitative Performance: As shown in the multi-event examples, SynCoS successfully manages transitions—such as a gorilla moving from playing a guitar to leaning back—without losing the character's visual identity or the scene's lighting consistency.
Discussion and Conclusion
SynCoS provides a robust, tuning-free solution for long-form video generation by treating global consistency as a refinement process rather than a standalone generation task. While the framework is computationally more intensive than simple fusion methods—taking approximately 55 minutes for a 30-second video on an NVIDIA H100—it offers superior visual stability and prompt fidelity. Future work will focus on optimizing the refinement stage to reduce inference time and exploring even longer video generation beyond the current 5x extension of base model capacities.
Summary updated