Introducing Paris 2.0 — Decentralized Trained Open-Weight Video Generation Model

TL;DR

Paris 2.0 is Bagel Labs’ video-generation follow-up to Paris (Paris: A Decentralized Trained Open-Weight Diffusion Model), and the first video diffusion model the team claims was pre-trained across heterogeneous GPU types distributed across geographic regions. The recipe inherits the original Paris design — partitioned data, expert diffusion models trained in complete isolation with zero gradient/parameter/activation synchronization, plus a small router selecting experts at inference — but extends it to video and to a deeper hardware heterogeneity (mixed GPU generations and vendors, multiple cloud regions). The headline result is a ~2.0× FVD improvement (561.04 → 279.01) over a monolithic baseline trained at matched total compute on low-resolution text-to-video. The page is a launch announcement; a full technical report is not linked from the blog post at filing time.

Key claims

Paris 2.0 is described as the first video generation model pre-trained across heterogeneous GPU types distributed across regions [§Announcement, ¶1].
The training pool spans GPU generations and vendors, not just multiple regions of one hardware family [§“Three things make Paris 2.0 unique”, point 1].
Training was distributed across geographically separated regions and clouds rather than running inside a single shared datacenter [§“Three things make Paris 2.0 unique”, point 2].
On low-resolution text-to-video, against a monolithic baseline at matched total compute budget, Paris 2.0 reports FVD 279.01 vs the baseline’s 561.04 — a ~2.0× reduction [§“Three things make Paris 2.0 unique”, point 3].
The blog post body does not state the resolution, dataset, number of experts, or per-expert parameter count for Paris 2.0; the launch defers those to a forthcoming technical report [§body — absence].

Method

The blog post is a short launch announcement and does not detail the architecture. By continuity with Paris: A Decentralized Trained Open-Weight Diffusion Model (same authors Villagra & Roy, plus new authors Ali Rouzbayani and Gin Jiang from Bagel Labs), the implied recipe is the same Distributed Diffusion Training framework: partition training data into semantically coherent clusters, train one expert diffusion model in complete isolation per cluster with no gradient, parameter, or activation exchange between experts, and post-hoc train a lightweight transformer router that selects experts during the reverse diffusion process. The novel claim relative to v1 is that the training pool is deeply heterogeneous along two axes simultaneously — GPU generation/vendor mix and geographic/cloud-region mix — which the zero-communication protocol is designed to make tractable. Quantitative configuration (number of experts, parameter count per expert, dataset, training budget in GPU-days) is not stated in the launch post.

Results

Low-resolution text-to-video, monolithic baseline vs Paris 2.0 at matched total compute: FVD 561.04 → 279.01, a ~2.0× improvement [§“Three things make Paris 2.0 unique”].
No resolution, dataset, or compute-budget number is given in the launch post; the v1 comparison axis (data and compute reduction relative to a non-Paris baseline) is not repeated for v2.
No qualitative samples or per-axis ablation (heterogeneous-hardware-only vs. cross-region-only vs. both) is reported in the launch text.

Why it’s interesting

This is the second filed datapoint from the zero-communication-training line and the first that targets video — the prior filed Paris (Paris: A Decentralized Trained Open-Weight Diffusion Model) was text-to-image only. Of the five systems currently filed under Distributed training parallelism that reduce communication overhead (veScale-FSDP RaggedShard, SpectraX MPMD, Fireworks 4D, KnapFormer compute-bags, VeOmni Async-Ulysses), all still assume periodic synchronization; Paris 1.0 was the only filed system that eliminated it, and Paris 2.0 extends that endpoint of the spectrum from images to video at matched compute. The reported 2× FVD improvement over a same-compute monolithic baseline is the inverse direction of the v1 claim (v1 was “comparable quality at 14–16× less compute”; v2 is “better quality at the same compute”). Whether that flips because video benefits more from data-partitioning than images, or because the v2 baseline is weaker, is not resolvable from the launch post — both readings remain open until the technical report lands. As an organizational signal it also reinforces Bagel Labs’ published roadmap, which had named video generation as the next target after Paris 1.0.