inferwire
/
AI·5 min read

Paris 2.0: Video Diffusion Trained Without Synchronized GPUs

Bagel Labs trained three 11B video experts in isolation and routed between them. Against one model with the same data and compute, FVD fell from 561.04 to 279.01.

TL;DR

  • Bagel Labs' Paris 2.0 trains three 11-billion-parameter video experts that never exchange gradients, parameters or activations, and routes between them at inference [^1].
  • Against one model trained on the same data with matched compute, it cut Fréchet Video Distance from 561.04 to 279.01 in a low-resolution test [^1].

Background

Diffusion models generate images and video by gradually turning noise into a picture. Large ones are usually trained as a single model whose parameters are updated together, which ties training to a cluster of co-located GPUs behind high-bandwidth interconnects [^1]. A decentralized diffusion model (DDM) instead splits the training data into clusters and trains a separate expert on each, with no communication between them. A small router then decides during generation which experts to use [^1].

What happened

Paris 2.0, posted to arXiv on May 25 by Ali Rouzbayani, Bidhan Roy, Marcos Villagra and Zhiying Jiang of Bagel Labs, extends that recipe from images to video [^1]. Its predecessor, Paris 1.0, was an open-weight image model built from eight experts of 129 million to 605 million parameters each, trained asynchronously across AWS, GCP, local clusters and Runpod in 120 A40 GPU-days [^2]. Video is harder, the authors write, because it burdens every expert with motion, longer temporal context and heavier latents, so it had been an open question whether the recipe could produce temporally coherent video [^1].

Each Paris 2.0 expert is an 11-billion-parameter FLUX-style MM-DiT, a multimodal diffusion transformer initialized from FLUX.1-dev image weights and pre-trained for video on a single data cluster [^1]. The router is a much smaller model of roughly 100 million parameters. At every denoising step it reads the current noisy video, the timestep and a compact text vector, then weights the top experts and combines their outputs [^1]. Because the experts exchange nothing during training, each could run on the cheapest hardware available, including spot instances and consumer hardware spread across clouds and regions, the paper says [^1]. It does not say which hardware Paris 2.0 itself was trained on.

The reported study is a first stage. Three experts, one per data cluster, faced a single 11-billion-parameter model trained on all three clusters together; each expert saw a third of the data with a third of the baseline's compute [^1]. Both were scored on 2,048 held-out clips at 256 × 256 pixels with identical sampling settings. The experts cut Fréchet Video Distance, a measure of how realistic generated videos look as a whole, from 561.04 to 279.01, and raised CLIP text-video similarity from 0.2032 to 0.2178 and the aesthetic score from 3.7950 to 3.9036 [^1]. Bagel Labs put the experts and router on Hugging Face behind a manually reviewed access request. The form says the weights are for research and evaluation and commercial use needs a written agreement, while the same page lists an MIT license [^3].

Why it matters

The obvious objection is size: the ensemble holds more parameters in total than the baseline, which the authors name as the only structural difference [^1]. They argue that routing, not raw capacity, carries the gain. With the router bypassed, alternating between two experts across denoising steps beat either expert alone on CLIP, aesthetic and warping scores, and 24 of 40 prompts preferred the switching schedule; one expert seemed to prefer high-noise steps, the other low-noise steps [^1]. One expert scored a CLIP similarity of 0.2175 on its own cluster's prompts against 0.1781 on generic prompts, so the experts do specialize [^1]. Paris 1.0 showed the same pattern for images: weighting all eight experts gave an FID of 47.89, worse than the single model's 29.64, while the top two experts reached 22.60 [^2].

The evidence is still narrow: three selected clusters, low resolution and automated metrics. The DDM also produced more motion, 0.712 against 0.555 pixels per frame, but the authors stress that this measures per-frame displacement, not temporal consistency [^1]. If the result holds at scale, capacity grows by training another expert rather than enlarging one synchronized run, while each sample still costs about one backbone's compute [^1]. The authors point to world models for physical AI as the next target, "trained on the distributed compute the world already has rather than the monolithic clusters only a few labs can assemble" [^1].

Practical example

Say Tomasz runs a five-person startup that wants its own video model for cooking tutorials, craft videos and product demos. A conventional training run would need a reserved block of GPUs behind fast interconnects, because every training step synchronizes across them [^1]. That is out of Tomasz's budget.

With the DDM recipe, the team first sorts its clip library into clusters, one per type of video. Each expert then trains on its own cluster, on whatever machines are cheap that week: spot instances in one cloud, a rented server in another. When a spot instance disappears, nothing else waits for it, because no expert depends on another's gradients; Paris 1.0's authors list this as no blocking from slow workers [^2]. The router trains separately, as a classifier that guesses which cluster a noisy clip came from [^1].

At generation time, a prompt for a stir-fry tutorial reaches the router at each denoising step, which weights the experts so that the cooking expert does most of the work. Each video still costs roughly one expert's compute [^1]. Adding a new category later means training another expert and updating the router, not rerunning one giant job [^1].

Tomasz should keep two limits in mind. Paris 2.0's evidence so far comes from a low-resolution study at 256 × 256 [^1]. And using Bagel Labs' own weights in a product needs the company's approval and a written agreement [^3].

Related gear

We recommend this book because it explains how neural networks are trained with gradient updates, including how large-scale training spreads that work across machines, the coordination Paris 2.0 does without.

AdvertisementAmazon

Deep Learning (Adaptive Computation and Machine Learning series)

Sources

  1. [1]arXiv — Paris 2.0: A Decentralized Diffusion Model for Video Generation
  2. [2]arXiv — Paris: A Decentralized Trained Open-Weight Diffusion Model
  3. [3]Hugging Face — bageldotcom/paris2
  4. [4]Hacker News — Paris 2.0: Video diffusion model trained on decentralized, heterogeneous GPUs