Bertrand Charpentier

Alibaba models
- replicate.com/wan-video/wa...
- replicate.com/wan-video/wa...
- replicate.com/prunaai/vace...
- replicate.com/prunaai/wan-...
- replicate.com/qwen/qwen-im...

HiDream models:
- replicate.com/prunaai/hidr...
- replicate.com/prunaai/hidr...
- replicate.com/prunaai/hidr...
- replicate.com/prunaai/hidr...

Flux models:
- replicate.com/prunaai/flux...
- replicate.com/prunaai/flux...
- replicate.com/prunaai/flux...

In the past months, we smashed all the state-of-the-art image and video models to make them x2-x10 faster!

We compressed All models from Black Forest Labs, HiDream, and Alibaba on @replicate.com 👇

We even created it the super high resolution Wan Image which can make cinematic images in seconds ;)

From Wan video to Wan Image: We built the fastest endpoint for generating 2K images!

- Accessible on @replicate.com: replicate.com/prunaai/wan-...
- Check details, examples, and benchmarks in our blog: www.pruna.ai/blog/wan-ima...
- Use Pruna AI to compress more AI models: docs.pruna.ai/en/stable/

GitHub - PrunaAI/awesome-ai-efficiency: A curated list of materials on AI efficiency A curated list of materials on AI efficiency. Contribute to PrunaAI/awesome-ai-efficiency development by creating an account on GitHub.

Happy to announce the "Awesome AI Efficiency" repo gathering 100+ curated materials on efficient AI

It is designed for a large audience and includes diverse materials.

It is an ongoing project. Do not hesitate to share your feedback/suggestions and star the repo! 🌟

github.com/PrunaAI/awes...

🚀 𝟭𝟬,𝟬𝟬𝟬 𝗠𝗼𝗱𝗲𝗹𝘀 𝗦𝗺𝗮𝘀𝗵𝗲𝗱 𝗼𝗻 @hf.co 🤗

(+2,500 in just the last two months 📈 all of you are keeping us busy!)

To celebrate, we gave our HF space a little makeover… and snuck in an Easter Egg 🥚

👀 Think you can find it? Drop a comment if you do!

Very clear example of how two (Bayesian) agents with wrong prior beliefs can lead to belief polarization even when seeing the same evidence! 👏

Super excited to share that the MatterGen code is now public on GitHub! github.com/microsoft/ma...

Big news for the AI community! Pruna AI is now free(mium)! 🎉

It takes a based model as input and returns a compressed model as output with various compression methods.

You can access via:
- Self-hosted w/ CLI (docs.pruna.ai/en/latest/se...)
- Cloud hosted w/ AWS: aws.amazon.com/marketplace/...).

Hope you will enjoy the paper and the code 🙂

Paper: arxiv.org/pdf/2402.15978
Code: github.com/fortuinlab/s...

In this paper 📄

- We train sparsifiable models via the Marginal Likelihood to make pruning easy after training ✅
- We construct parameter- & unit-wise prior that fit unstructured and structured pruning ✅
- We propose the cheap Optimal Posterior Damage score outperforming other pruning scores ✅

“Shaving Weights with Occam’s Razor: Bayesian Sparsification for Neural Networks using the Marginal Likelihood” is at #NeurIPS2024!

This would not have been possible without Rayen Dahri* who lead the project, Alex Immer, Stephan Günnemann, and @vincefort.bsky.social !

1/3

This was a super fun collaboration, led by my master's student Rayen Dhahri with Alex Immer,
@bertrand-sharp.bsky.social, and Stephan Günnemann

Hope you will enjoy the paper and the overview 🙂

Paper: openreview.net/pdf/b84db14f...
Overview: www.cs.cit.tum.de/daml/expecte...

4/4

It is part of a long research on structure learning :)

- "Differentiable DAG Sampling", ICML 2022
- "End-to-end Learning of Probabilistic Hierarchies on Graphs", ICLR 2021
- "Tree Sampling Divergence", IJCAI 2019
- "Hierarchical graph clustering using node pair sampling" MLG, KDD 2018

3/4

In this paper 📄

- We optimize the expected objectives over probabilistic model on hierarchies. ✅
- We differentiably sample any hierarchy (not only binary ones!). ✅
- We show theoretical consistency and achieve SOTA results for hierarchical learning. ✅

2/4

“Expected Hierarchical Clustering” is at #NeurIPS2024!

This would not have been possible without Marcel Kollovieh* who lead the project, @danielzuegner.bsky.social and Stephan Günnemann !

1/4

🚀 𝗟𝗼𝗼𝗸𝗶𝗻𝗴 𝗔𝗵𝗲𝗮𝗱
Our journey continues as we grow our team of 𝘁𝗼𝗽 𝘁𝗮𝗹𝗲𝗻𝘁𝘀 𝗮𝗰𝗿𝗼𝘀𝘀 𝗘𝘂𝗿𝗼𝗽𝗲 🇪🇺, with hubs in Germany and France 🇩🇪🤝🇫🇷. Together, we’re enabling cutting-edge AI compression and deployment without compromising real-world performance.

🌟 𝗟𝗼𝗼𝗸𝗶𝗻𝗴 𝗕𝗮𝗰𝗸
What started as an idea in my PhD, became an incredible journey. Bringing together my brilliant co-founders -S. Günnemann, J. Rachwan, R. Nait Mazi- was a pivotal step for Pruna AI. In 1 year, we smashed thousands of AI models, making them 𝗰𝗵𝗲𝗮𝗽𝗲𝗿💲, 𝗳𝗮𝘀𝘁𝗲𝗿⚡, 𝘀𝗺𝗮𝗹𝗹𝗲𝗿💾, and 𝗴𝗿𝗲𝗲𝗻𝗲𝗿🌳.

@prunaai.bsky.social has raised $𝟲.𝟱𝗠 to make AI accessible to everyone 👥 and sustainable for the planet 🌍!

#EfficientAI #AccessibleAI #SustainableAI #TechForGood

I would be happy to be part of it :)

That looks great! I would love to be in :)