Stephan Günnemann

Wan 2.1 might be the best open-source text-to-image model, and everyone is sleeping on it. The one drawback is Wan's slow inference speed, so we applied a series of optimizations to bring it down to just 3s for 2 MP images. You can try it on @replicate: replicate.com/prunaai/wan-im…

🧑‍🏫 AI Efficiency Fundamentals - Week 4: Quantization We see quantization everywhere, but do you know the difference between static and dynamic quantization? Even if you do, these slides are great for you. At Pruna, we want to educate about efficient AI, so our lead researcher…

Yesterday we launched wan-image on @replicate and it can generate amazing animal cinematic pictures

From Wan video to Wan Image: We built the fastest endpoint for generating 2K images! - Accessible on @replicate : lnkd.in/eqsBR2Kx - Check details, examples, and benchmarks in our blog: lnkd.in/eXcAbqjM - Use Pruna AI to compress more AI models:…

📷 Introducing Wan Image – the fastest endpoint for generating beautiful 2K images! From Wan Video, we built Wan Image which generates stunning 2K images in just 3.4 seconds on a single H100 📷 Try it on @replicate: replicate.com/prunaai/wan-im… Read our blog for details, examples,…

We're pleased to work with Pruna to bring you a new and fast image model. It can generate 2 megapixel images in 3.4 seconds on a single H100 replicate.com/prunaai/wan-im… This model is based on the original Wan 2.1 video model, which Pruna have compressed, optimised and pruned.

How private is DP-SGD for self-supervised training on sequences? Our #ICML2025 spotlight shows that it can be very private—if you parameterize it right! 📜arxiv.org/abs/2502.02410 #icml Joint work w/ M. Dalirrooyfard, J. Guzelkabaagac, A. Schneider, Y. Nevmyvaka, @guennemann 1/6

We’re thrilled to share that the first in-person LoG conference is officially happening December 10–12, 2025 at Arizona State University logconference.org Important Deadlines: Abstract: Aug 22 Submission: Aug 29 Reviews: Sept 3–27 Rebuttal: Oct 1–15 Notifications: Oct 20

🇪🇺✈️🇺🇸 In SF next week. Optimizing AI models & handing out croissants to @ycombinator startups haunted by Soham. DM before the croissants vanish 🥐 @PrunaAI

🫟 Open-weights @bfl_ml FLUX.1 Kontext [dev] is now open-source! It allows to perform image-to-image generation with state-of-the-art quality :) However, it takes ~14.4 seconds for each generation on one H100. When we learned about this, we were in our offsite to chill together…

💎 Congrats @ComfyUI, a shiny new logo, deserves a blazing fast new Pruna integration, resulting in 𝟲𝗫 𝗳𝗮𝘀𝘁𝗲𝗿 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀! 🏎️ We created 𝗰𝘂𝘀𝘁𝗼𝗺 𝗣𝗿𝘂𝗻𝗮 𝗻𝗼𝗱𝗲𝘀 𝗮𝗻𝗱 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝗲𝗱 𝘁𝗵𝗲𝗺. We proved we accelerate inference by 6X, reducing…

📣 Want to speed up your inference time? Join us to learn how to optimize your AI models with @PrunaAI and deploy on high-performance serverless GPUs in seconds with @gokoyeb 🚀 🔥 Faster deployments ⚙️ Optimized inference ✅ Deploy in seconds app.livestorm.co/pruna-ai/pruna…

✨ Pruna now supports saving and loading models with @huggingface - x.com/huggingface -! 🧃 Juicy workflow! • Load pre-trained model • Smash and optimize the model • Save it to Hugging Face • Load the optimized model again 🤺 Production ready optimization in one…

How do LLMs navigate refusal? Our new @ICMLConf paper introduces a gradient-based approach & Representational Independence to map this complex internal geometry. 🚨 New Research Thread! 🚨 The Geometry of Refusal in Large Language Models By @guennemann's lab & @GoogleAI. 🧵👇

📢Apply now for the relAI MSc program! Get cross-sectional AI training & a scholarship while pursuing your Master's at @TU_Muenchen or @LMU_Muenchen. Deadline: June 17, 2025. Link: zuseschoolrelai.de/application/#M… @DAAD_Germany @GittaKutyniok @guennemann @baiosphere_AI

🔥 𝗬𝗼𝘂𝗿 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗹𝗼𝘄𝗲𝗿 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, 𝗵𝗶𝗴𝗵𝗲𝗿 𝘁𝗵𝗿𝗼𝘂𝗴𝗵𝗽𝘂𝘁, and 𝗿𝗲𝗱𝘂𝗰𝗲𝗱 𝗺𝗲𝗺𝗼𝗿𝘆 𝘂𝘀𝗮𝗴𝗲 for any production-scale AI deployments! Last week we created FLUX-juiced, 2.6x faster and much more efficient, but…

Today, we are excited to announce our partnership with @PrunaAI! 🎉 Optimize your AI models with Pruna and deploy them on our high-performance serverless GPUs in seconds 🚀 Read the announcement to get started and learn more 👇 koyeb.com/blog/achieve-5…

One click. 5x faster inference on serverless GPUs. Pruna x Koyeb partnership = 🚀🚀🚀 ➡️ 5x faster speeds ➡️ Infra cost savings ➡️ Zero scaling pain Optimized models like Flux, Llama, Wan, Stable Diffusion… ready to fly on Koyeb’s H100s, A100s, L40s. One click deploy. No…

If you are attending #ICLR2025 and are interested in electronic structure modelling / quantum chemistry come by our poster on learnable non-local XC-functionals to discuss with @n_gao96 and me. 🗓️ Today | 3:00 pm– 5:30 pm 📍Hall 3 | Poster #3

Thrilled to announce that we just presented „MAGNet: Motif-Agnostic Generation of Molecules from Scaffolds“ at #ICLR2025 🧲 @j_m_sommer @Pseudomanifold @fabian_theis @guennemann For those who couldn’t make it to our spotlight: openreview.net/forum?id=5FXKg…