Paris: A Decentralized Trained Open-Weight Diffusion Model

Imagine a world where creating art isn't just about a single genius sitting in a studio, but about a massive, global team of artists working together without ever needing to talk to each other. This is the heart of a field called "diffusion modeling," a type of artificial intelligence that learns to create images by starting with a chaotic cloud of static noise and slowly, step-by-step, refining it into a clear picture, much like a sculptor chipping away stone to reveal a statue. For years, the biggest hurdle in this field has been the need for a "central command center." To train these powerful AI artists, researchers usually had to gather thousands of super-fast computers in one place, connected by expensive, high-speed cables, all working in perfect lockstep. If one computer lagged, the whole team had to wait. This meant only the richest, most well-funded institutions could build the best image generators, leaving everyone else out of the game.

Enter "Paris," a new project by Bagel Labs that asks a bold question: What if we didn't need that central command center at all? What if we could train a world-class AI artist by letting many smaller teams work in total isolation, on their own computers, in different countries, and only bring their work together at the very end? The paper presents Paris as the first publicly released image generator trained entirely this way, proving that you don't need a supercomputer cluster to make beautiful pictures. Instead, it uses a clever trick where the AI is split into a team of specialists, each learning a different style, and a smart "traffic controller" that decides which specialist to listen to when you ask for a picture.

The Big Idea: A Team of Solitary Geniuses

Usually, training a massive AI is like trying to get a choir of 1,000 singers to harmonize perfectly. They all have to stand in the same room, listen to a conductor, and adjust their voices instantly if someone is off-key. If the room is too big or the microphones are slow, the whole song falls apart. This is what happens with traditional AI training: thousands of graphics cards (GPUs) constantly shout their progress to each other, waiting for everyone to catch up before moving to the next step. It's expensive, requires special high-speed internet, and is impossible to do with a mix of old and new computers.

Paris flips the script. Instead of one giant choir, imagine a group of eight soloists, each locked in their own soundproof room. They never talk to each other, never share their sheet music, and never wait for the others to finish.

• The Soloists (The Experts): The Paris model is made of eight smaller AI "experts." Each one is trained on a different slice of the internet's image library. One might only learn about sunsets, another about dogs, and another about architecture. They train in complete isolation, on whatever hardware is available, even if one is on a fast server in the US and another is on a slower machine in Europe.
• The Traffic Controller (The Router): Since these experts never talk during training, how do we know which one to use when you type "a golden retriever running through a meadow"? That's where the "router" comes in. It's a tiny, lightweight AI that acts like a smart traffic cop. When you give it a prompt, it looks at the messy, noisy starting point of the image and quickly decides, "Hey, Expert 3 is the best at dogs, let's use them!" or maybe, "Actually, this scene needs a mix of Expert 3 and Expert 5."

How It Works: The Magic of "No Talking"

The genius of Paris lies in how it handles the training data. The researchers took a huge dataset of 11 million images and used a smart tool (called DINOv2) to sort them into eight distinct piles based on what they look like. Then, they sent each pile to a different expert.

• No Synchronization: In normal AI training, the computers have to stop and sync up their "gradients" (which are like notes on how to improve) constantly. Paris experts don't do this. They just train, train, train, on their own time, at their own speed. One might finish 100,000 steps while another is only at 90,000. It doesn't matter.
• The Router's Job: The router is trained separately. It learns to look at a noisy, blurry image and guess which expert is the best fit for that specific noise pattern. It's like a librarian who knows exactly which bookshelf to pull from based on the faint smell of the book you're holding.

What They Found: Less Data, Less Power, Same Quality

The team tested Paris against the old way of doing things and against a previous decentralized attempt. The results were surprisingly efficient:

• Resource Savings: Paris managed to train a high-quality model using 14 times less training data (11 million images instead of 154 million) and 16 times fewer GPU-days (the amount of computer time needed) compared to the previous best decentralized method.
• The "Top-2" Surprise: When they tested how to combine the experts, they found something interesting. Using just the single "best" expert (Top-1) was good, but using the top two experts and mixing their results together (Top-2) actually produced the best pictures.
• The "Full Team" Mistake: Interestingly, trying to use all eight experts at once (the Full Ensemble) made the pictures worse. It turns out that having too many voices trying to sing at once creates noise. The paper suggests that being selective is key; you don't want the whole choir, you just want the right soloists.

Why This Matters

The paper concludes that you don't need a billion-dollar data center to build a world-class image generator. By letting computers work independently and only coordinating them at the very end, Paris proves that high-quality AI can be built on "heterogeneous hardware"—meaning a mix of different, cheaper, and geographically scattered computers. This opens the door for researchers and smaller companies to train powerful models without needing the massive, synchronized infrastructure that has been the gatekeeper of the field until now.

In short, Paris shows that you can build a masterpiece not by forcing everyone to march in step, but by letting everyone find their own rhythm and then having a smart conductor bring them together just in time for the final bow.