ByteDance also open-sourced BAGEL, a 7B multimodal model.
Robert Scoble#447
We're releasing Paris 2.0, which, to our knowledge, is the world's first decentralized trained video generation model.
We benchmarked it against a monolithic model trained on the same data and compute budget, and Paris 2.0 outperformed the monolithic by ~2x on FVD benchmark.
Many users praised Bagel Labs' Paris 2.0 release as an inspiring first decentralized video generation model because its open-source approach and training method represent a major advance over centralized systems.
No Digg Deeper questions have been answered for this story yet.
ByteDance just open-sourced one of the most capable multimodal models out there.
BAGEL does image generation, editing, style transfer, and visual understanding - all in a single 7B parameter model. Apache 2.0 licensed!
One model. No switching between specialized tools. Amazing
We're releasing Paris 2.0, which, to our knowledge, is the world's first decentralized trained video generation model.
We benchmarked it against a monolithic model trained on the same data and compute budget, and Paris 2.0 outperformed the monolithic by ~2x on FVD benchmark.
We're selectively releasing the Paris 2.0 weights and partnering with researchers and teams interested in diffusion-based video models, world models, and embodied agents. The model is on Hugging Face: https://huggingface.co/bageldotcom/paris2
Video diffusion backbones increasingly anchor the world models behind physical AI. Our early internal results show the Paris 2.0 recipe carries over to them very well. We will share more on that in future work.
For the nerds out there:
The way this model was trained was pretty cool, and probably, the first time this has been done to train a video generation model.
If you want to train a model, you rent a huge GPU cluster and do it there.
But this model was trained differently:
The team that built this model trained different "experts" using separate GPU clusters and data. These GPUs were all different, and didn't need any communication between them.
So instead of allocating 100 interconnected GPUs to train the model, they used 3 GPUs here, 5 there, 4 more over there, etc.
After training, they added a smart router on top of all of the trained experts. This router's job is to take an inference request, and route it to the appropriate experts.
In other words, instead of training a single model, they trained multiple smaller models that work together at inference time.
And the results were really good!
Here is the link to their paper: https://arxiv.org/pdf/2605.26064. It's a very good read.
We're releasing Paris 2.0, which, to our knowledge, is the world's first decentralized trained video generation model.
We benchmarked it against a monolithic model trained on the same data and compute budget, and Paris 2.0 outperformed the monolithic by ~2x on FVD benchmark.
Paris 2.0 builds on Paris 1.0, which proved that image generation can be trained on a geographically distributed, heterogeneous pool of GPUs. Temporally coherent video under the same settings remained an open question. Paris 2.0 closes it.
Read the technical report on arxiv: https://arxiv.org/abs/2605.26064
Paris 2.0 is a Decentralized Diffusion Model (DDM). A DDM is an ensemble of independent diffusion models trained in isolation, each on its own slice of the data, and exchanges no gradients, parameters, or activations between them.
During inference, a lightweight router selects a subset of experts at each denoising step.
@kimmonismus Your automated AI poster made a mistake. Bytedance doesn't own Bagel Labs. That's a different project named Bagel under Bytedance
Video diffusion backbones increasingly anchor the world models behind physical AI. Our early internal results show the Paris 2.0 recipe carries over to them very well. We will share more on that in future work.
With Paris 2.0, our goal was only to match a monolithic model's benchmarks while training on a distributed, heterogeneous pool of GPUs. It beat them instead.
FVD dropped from 561.04 to 279.01, and CLIP text-video alignment and aesthetic score increased. To our knowledge, the DDM is the first distributed training architecture to surpass its monolithic counterpart under matched data and compute.
Congrats to my team members @roze12321 @mdvillagra24 @ZhiyingJ !
@bidhan Congrats Bidhan :). Amazing work!!!!! Can’t wait to try this out
@kimmonismus 1) you're confusing bytedance's "bagel" model with "paris" from "bagel labs". 2) skeptical of Paris. they're gating it, and not for commercial use. they publish only a few frames and a few look bad. If it were good, they wouldn't do that.
@Omarboucher I see. So the bad news is we won't be able to run the model on that machine right now. But the good news is Bagel Labs is working on an inference engine for DDM models actively, which will make this possible very soon. Please stay tuned 🫡
@bidhan Very interesting direction. I don't think I've seen anything like it.
@bidhan Let’s go! Niceee
@bidhan at first i thought this was only trained on data from the city of Paris 🤣
@bidhan Claude who? Paris is eating now 😎
@Kyrannio Thank you for your support Kiri!
@supremebeme wait, that's a good idea for the next model
@bidhan This is truly inspiring! The development of Paris 2.0, the world's first decentralized trained video generation model, is a monumental achievement.
@bidhan Congrats 👏
@MehakdeepK81 Thank you! We're trying hard to make open intelligence work.
