Paris 2.0: A Decentralized Diffusion Model for Video Generation

2026-05-25

AI Summary

This work addresses the challenge of generating temporally coherent videos under decentralized training conditions by proposing the first decentralized diffusion model (DDM) capable of high-quality text-to-video synthesis. The approach integrates low-resolution video diffusion pretraining with a distributed collaborative optimization strategy, enabling model training without reliance on centralized GPU clusters. Experimental results demonstrate that, under identical data and total compute budgets, the proposed method significantly reduces the Fréchet Video Distance (FVD) from 561.04 to 279.01—nearly a two-fold improvement—while also outperforming centralized baselines in both CLIP-based text-video alignment and aesthetic quality scores. This study thus achieves, for the first time, temporally consistent video generation within a decentralized architecture.

Abstract

We present Paris 2.0, the first video generation model pre-trained through decentralized computation. Its training recipe builds upon Paris 1.0 (arXiv:2510.03434), the first ever open-weight Decentralized Diffusion Model (DDM), which showed that image generation can be trained without a monolithic GPU cluster. However, temporally coherent video generation had remained an open problem under decentralized training, and Paris 2.0 closes it. In low-resolution text-to-video training, against a monolithic model trained on the same data under a matched total compute budget, Paris 2.0 cuts Frechet Video Distance (FVD) from 561.04 to 279.01, a ~2.0x improvement, and lifts CLIP text-video similarity and aesthetic score.

Problem

• decentralized training
• video generation
• temporal coherence
• diffusion model

Innovation

• Decentralized Diffusion Model
• Video Generation
• Temporal Coherence
• Frechet Video Distance
• Text-to-Video