Skip to content
Bud's Wiki

Paris 2.0 — decentralized trained video generation model from Bagel Labs (bidhan announcement)

TL;DR

Section titled “TL;DR”

Bidhan Roy (Bagel Labs CEO) announces Paris 2.0, the follow-up to the original Paris decentralized-trained image diffusion model (Paris: A Decentralized Trained Open-Weight Diffusion Model). The new release extends the zero-communication multi-expert recipe from text-to-image to text-to-video, and reports a ~2× FVD improvement over a monolithic baseline trained on the same data + compute budget. No technical report or weights link in the announcement tweet itself — only a video sample and the headline FVD claim. The artifact is the announcement; quantitative details beyond the headline ratio are not yet public.

Key claims

Section titled “Key claims”
  • Paris 2.0 is positioned as the world’s first decentralized-trained video generation model, extending the original Paris recipe (8 isolated DiT experts + post-hoc router) from images to video [tweet body].
  • A head-to-head comparison against a monolithic model trained on the same data and compute budget reports ~2× better FVD for Paris 2.0 [tweet body].
  • No architecture details, expert count, training-data scale, or compute-budget numbers are disclosed in the announcement [tweet body].

Method

Section titled “Method”

Not described in the announcement. The original Paris (Paris: A Decentralized Trained Open-Weight Diffusion Model §2) trained 8 DiT-XL/2 image-diffusion experts in complete isolation on DINOv2-clustered LAION-Aesthetic subsets, with no gradient / parameter / activation exchange, and a separately-trained DiT-B router selecting experts during the reverse-diffusion process. Paris 2.0 presumably extends this to a video DiT backbone with some video-tokenization layer (latent video, VAE, or otherwise) and an equivalent router strategy, but the tweet does not confirm any of this.

Results

Section titled “Results”
  • FVD improvement: ~2× over a monolithic baseline at iso-data, iso-compute [tweet body]. The baseline architecture, training-data corpus, FVD evaluation set, and absolute FVD numbers are not disclosed in the announcement.
  • 68.7K views and 55 replies on the announcement tweet at fetch time, indicating the release reached the broader ML Twitter audience but the technical details are pending [tweet metadata].

Why it’s interesting

Section titled “Why it’s interesting”

This is the second filed artifact from Bagel Labs in the Distributed training parallelism zero-communication line, after Paris: A Decentralized Trained Open-Weight Diffusion Model established the recipe on image diffusion. The video extension is the natural next step: video DiTs are exactly the workload most punished by all-reduce overhead because activation tensors are an order of magnitude larger than image latents, so the per-expert-isolation design should pay back more here than it did on images.

A 2× FVD gap over a monolithic iso-compute baseline — if it holds up on a public benchmark — would put Paris 2.0 in a different regime than the image release, which reported a 1.27× worse FID than centralized DDM (at 14× less data and 16× less compute). It would also be a stronger result than any other filed paper under MoE Routing Design for video, where joint MoE training with load-balancing losses (HunyuanVideo 1.5 Technical Report, Wan2.2: World's First Open-Source MoE Video Generation Model (launch announcement)) is the standard approach. Caveat: until the FVD evaluation set and monolithic baseline are public, this is a vendor claim, not a third-party comparison.

See also

Section titled “See also”