Paris: A Decentralized Trained Open-Weight Diffusion Model
Zhiying Jiang, Raihan Seraj, Marcos Villagra, Bidhan Roy
TL;DR
Paris shows that high-quality diffusion models can be trained without gradient synchronization by using a decentralized ensemble of $K=8$ expert Diffusion Transformers trained on partitioned data. A DINOv2-based semantic clustering assigns data to clusters, while a lightweight DiTRouter selects and combines expert predictions at inference, enabling zero inter-expert communication during training. The approach achieves substantial resource savings—approximately $11\mathrm{M}$ training images and $\sim 120$ GPU-days for DiT-XL/2, versus $158\mathrm{M}$ images and $\sim 1176$ GPU-days for a centralized baseline—while maintaining competitive FID performance (e.g., $\mathrm{FID}=12.45$ on LAION-Aesthetic vs $9.84$ for the baseline) with a nuanced trade-off depending on the inference strategy. The work provides practical, open-source patterns for distributed generative modeling on heterogeneous hardware, enabling research and commercial use without centralized compute infrastructure.
Abstract
We present Paris, the first publicly released diffusion model pre-trained entirely through decentralized computation. Paris demonstrates that high-quality text-to-image generation can be achieved without centrally coordinated infrastructure. Paris is open for research and commercial use. Paris required implementing our Distributed Diffusion Training framework from scratch. The model consists of 8 expert diffusion models (129M-605M parameters each) trained in complete isolation with no gradient, parameter, or intermediate activation synchronization. Rather than requiring synchronized gradient updates across thousands of GPUs, we partition data into semantically coherent clusters where each expert independently optimizes its subset while collectively approximating the full distribution. A lightweight transformer router dynamically selects appropriate experts at inference, achieving generation quality comparable to centrally coordinated baselines. Eliminating synchronization enables training on heterogeneous hardware without specialized interconnects. Empirical validation confirms that Paris's decentralized training maintains generation quality while removing the dedicated GPU cluster requirement for large-scale diffusion models. Paris achieves this using 14$\times$ less training data and 16$\times$ less compute than the prior decentralized baseline.