🤖 machine learning

Paris 2.0: A Decentralized Diffusion Model for Video Generation

Paris 2.0 is the first decentralized diffusion model capable of training temporally coherent video generation, achieving a roughly 2.0x improvement in Frechet Video Distance compared to monolithic models under a matched compute budget.

Original authors: Ali Rouzbayani, Bidhan Roy, Marcos Villagra, Zhiying Jiang

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ali Rouzbayani, Bidhan Roy, Marcos Villagra, Zhiying Jiang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a robot how to draw moving pictures (videos) based on a description you give it. Usually, to do this, you need a massive, super-expensive computer room filled with thousands of powerful graphics cards all working together in perfect lockstep. If one card stops, the whole team stops. This is the "monolithic" way of doing things.

Paris 2.0 is a new way of doing this that doesn't need that giant, expensive room. Instead, it uses a "decentralized" approach, like a team of freelancers working from different coffee shops around the world, all connected by a smart manager.

Here is how the paper explains it, broken down into simple concepts:

1. The Problem: The "Big Boss" vs. The "Team"

The Old Way (Monolithic): Imagine a single, giant brain trying to learn everything about video at once. It has to be trained on one giant supercomputer. It's expensive, hard to build, and if the internet connection to that one computer drops, the training stops.
The New Way (Paris 2.0): Imagine hiring three different artists (called Experts). Each artist works in their own small studio on their own computer. They don't talk to each other while they are learning. They just practice on their own specific pile of videos.
- The Magic: Because they don't have to wait for each other to finish a step, they can use cheaper, scattered computers (even ones people rent by the hour) to train.

2. How It Works: The "Smart Manager"

When you ask the system to make a video (like "a woman with blond hair speaking"), the system doesn't just pick one artist. It uses a Router (the Smart Manager).

The Process:
1. You type your prompt.
2. The video is built step-by-step, like peeling an onion layer by layer to reveal the picture.
3. At every single step of peeling the onion, the Router looks at the current messy image and asks: "Who is best at fixing this specific part?"
4. The Router might say, "Expert A is great at the beginning of the video, but Expert B is better at the end."
5. It instantly picks the right expert to do the next step.

The Analogy: Think of making a movie. In the old way, one director had to direct every single scene, from the explosion to the quiet conversation. In Paris 2.0, you have a director who knows exactly which specialist to call for each scene. One specialist is great at action, another at emotions. The "Router" switches between them instantly as the movie plays.

3. The Results: Better Quality, Less Cost

The paper tested this new team of experts against the old "giant brain" model. They gave both the exact same amount of computer power and the exact same data to learn from.

The Score: The new decentralized team (Paris 2.0) made videos that looked much more realistic and matched the text prompts better.
- They cut a major error score (FVD) in half (from 561 down to 279).
- The videos looked more beautiful and followed the instructions more closely.
The Surprise: The paper found that the "team" approach actually worked better than the giant single model, even though the total amount of computing power used was the same.

4. Why It Matters (According to the Paper)

The paper claims this proves that you don't need a massive, centralized supercomputer to train advanced video AI.

Specialization: The different "experts" naturally learned to be good at different types of videos (like different camera movements or scenes) because they trained on different chunks of data.
Scalability: If you want to make the AI smarter, you don't need to build a bigger supercomputer. You just hire another "expert" (train another model) and let the Router learn to use them.

In short: Paris 2.0 shows that by splitting the work among many smaller, independent models and using a smart system to switch between them, we can create high-quality video AI without needing the world's most expensive computers. It turns video generation from a "one giant brain" problem into a "cooperative team" problem.

Technical Summary: Paris 2.0 – A Decentralized Diffusion Model for Video Generation

1. Problem Statement

Current video generation models rely on a monolithic training paradigm, where a single model is trained on a broad mixture of video datasets using a homogeneous cluster of high-bandwidth, co-located GPUs. This approach creates significant infrastructure constraints, as parameter updates require synchronization across tightly coupled hardware. While Decentralized Diffusion Models (DDMs) have successfully demonstrated that image generation can be trained without monolithic clusters (as shown in Paris 1.0 [2]), extending this to temporally coherent video generation remained an open problem. Video generation imposes heavier burdens on experts, including motion modeling, longer temporal contexts, and substantially larger latent spaces, raising the question of whether decentralized, asynchronous training could yield high-quality, coherent video.

2. Methodology

Core Architecture

Paris 2.0 is an ensemble of independent diffusion experts combined at inference by a learned router. The system consists of three primary components:

Shared Preprocessing Stack: Videos are encoded into cached causal HunyuanVAE latents. Prompts are encoded using T5-v1.1-XXL and CLIP-ViT-L/14. This stack is decoupled from the training forward path to reduce computational overhead.
Expert Pool: The model utilizes three 11B-parameter FLUX-style MM-DiT (Multimodal Diffusion Transformer) experts. Each expert is trained from scratch on a distinct data cluster using a flow-matching velocity objective. Crucially, experts exchange no gradients, parameters, or activations during training; they are optimized in isolation on heterogeneous hardware (including spot instances and consumer hardware).
Router: A lightweight DiT-B model (~100M parameters) performs routing at every denoising step. It consumes the current noisy latent, diffusion timestep, pooled CLIP text vectors, and an optional DINOv2 first-frame feature. It outputs routing weights to select the top- $K$ experts. Unlike the experts, the router does not access full T5 embeddings, relying instead on the pooled CLIP vector.

Training and Inference

Training: Each expert is confined to its own data cluster. The router is trained independently as a supervised cluster classifier, predicting the source-cluster label of a noisy latent paired with its CLIP text feature.
Inference: At each denoising step, the router reads the current state and selects experts to evaluate the video velocity field. The weighted sum of these experts' velocity fields drives an ODE solver to the next latent. This design ensures that per-sample compute remains on the order of a single backbone, while total capacity scales horizontally by adding more experts.

3. Key Contributions

First Decentralized Video Model: Paris 2.0 is the first video generation model pre-trained entirely through decentralized computation, proving that temporally coherent video can be generated without monolithic GPU clusters.
Decoupled Compute for Video: The paper extends the decoupled-compute recipe from images to video, demonstrating that distinct data clusters exhibiting different motion patterns and scene dynamics can be exploited by specialized experts without forcing every sample to pay for every expert.
Expert Specialization: The study provides evidence that experts trained in isolation genuinely specialize in their assigned data clusters rather than collapsing to a marginal distribution, a precondition for effective routing.
Routing vs. Ensembling: Through ablation studies, the authors demonstrate that performance gains stem from the routing mechanism across specialized experts rather than simple ensembling or increased parameter count alone.

4. Experimental Results

The authors evaluated Paris 2.0 (Stage 1, three-expert DDM) against a monolithic 11B MM-DiT trained on the union of the same data clusters under an iso-FLOP and iso-data budget (each expert sees 1/3 of the data and compute of the monolithic model).

Quantitative Improvements (Low-resolution 256×256 text-to-video):

Fréchet Video Distance (FVD): Reduced from 561.04 (Monolithic) to 279.01 (DDM), representing a ~2.0× improvement in distributional realism.
CLIP Text-Video Similarity: Increased from 0.2032 to 0.2178.
Aesthetic Score: Improved from 3.7950 to 3.9036.
Motion: The DDM showed higher per-frame displacement magnitude (0.712 vs 0.555), though the authors note this is a descriptive measure of magnitude, not temporal consistency.

Ablation Findings:

Switching Schedule: An alternating manual schedule across two experts outperformed single experts on all metrics, suggesting directional specialization across denoising time (e.g., one expert preferring high-noise steps, another low-noise).
Specialization Gap: An expert trained on a specific cluster achieved a CLIP score of 0.2175 on its own cluster's prompts versus 0.1781 on a generic set, confirming genuine specialization.

5. Significance and Claims

The paper claims that Paris 2.0 closes the open question of whether decentralized training can produce temporally coherent video. The primary significance lies in demonstrating that expert capacity trained in isolation not only survives the transition to video but outperforms a single backbone trained on the same data.

The authors position this work as a step toward physical AI world models. They argue that foundation world models must cover environments far beyond the capacity of any single training cluster. By specializing experts per environment and routing across them, Paris 2.0 offers a pathway to train world models on the distributed compute the world already has, rather than relying on the monolithic clusters accessible only to a few labs. The results suggest that the DDM recipe is robust enough to scale from images to complex, motion-heavy video generation.