Paris

Paris demonstrates that high quality text to image generation does not require a centralized GPU cluster. The team trains a set of expert diffusion models independently, with no gradient, parameter, or activation synchronization during pretraining. This zero communication recipe lets volunteers or heterogeneous nodes contribute compute, avoids dependencies on NVLink or InfiniBand, and scales with simple orchestration. In experiments, the approach reports large efficiency gains versus conventional pipelines while maintaining competitive visual quality, and the weights are released for open use. In practice, Paris is a practical foundation for creative tools and research that want modern diffusion images and a training method that works across scattered hardware.