Stephan Günnemann
@guennemann
Followers
4K
Following
451
Media
3
Statuses
259
Professor for Data Analytics and Machine Learning @ TU Munich
Joined December 2019
Our paper “Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow” has been accepted at #NeurIPS2025! @ale__palmaa @guerrantif @fabian_theis @guennemann (@TU_Muenchen × @HelmholtzMunich × @MunichCenterML × @MdsiTum ) 🧵
3
12
22
Important insight to increase the reliability of LLMs.
Can we actually tell when LLMs know or don’t know? For questions with a single answer, that works. But once ambiguity enters - several answers are correct - current methods collapse, confusing model with data uncertainty. w/ @dfuchsgruber @TomWollschlager @guennemann [1/4]🧵
0
1
5
Open-source Diffusion LLMs easily break GPT-5! In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access. @TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn 🧵
1
16
19
🎬 Introducing 𝗪𝗮𝗻 𝟮.𝟮 𝗔𝗻𝗶𝗺𝗮𝘁𝗲: Copy motion from any video and bring your images to life, now with cinematic quality and MoE architecture! • Accessible on https://t.co/bm7ryqASpc • Use Pruna AI to compress more AI models: https://t.co/MCXKk7LJTb
0
3
9
You can now run Hunyuan-Image-3.0 on Replicate https://t.co/C74NzNowbg The #1 model in Text-to-Image LMArena Create images in under 30 seconds Another @PrunaAI collab to deliver the fastest speeds possible
7
14
80
Faster inference with Pruna x @replicate • 𝗦𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝘁 𝘀𝗽𝗲𝗲𝗱𝘂𝗽𝘀: reduce inference from seconds to milliseconds • 𝗟𝗮𝘆𝗲𝗿𝗲𝗱 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻: quantization, pruning, compilation, and caching combined • 𝗘𝗮𝘀𝘆 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁: optimized
1
2
11
You can now run Ovi by @character_ai Simultaneously generate both audio & video in under 40 seconds https://t.co/r3GWDPNxwz Another @PrunaAI collab to deliver the fastest speeds possible
11
7
53
📸 Introducing Hunyuan Image 2.1: Generate high resolution 2K images with text prompt support for just ~$0.02 per image! • Accessible on https://t.co/kJfxYUm0O0 • Use Pruna AI to compress more AI models: https://t.co/MCXKk7LJTb
1
1
7
Come and listen to @Bertrand_Charp share his talk: “How to make your AI models faster, smaller, cheaper, greener?” at the @aiDotEngineer 📍@joinstationf, Founder’s Cafe at 10:30 PM CEST Hope to see you there!
0
3
6
The hunt for increased WL expressivity has led to many new GNNs but limited real-world success. So what are we missing? Can we find a better objective? We answer these questions in our new paper: https://t.co/jrNtzM9mdY Joint work /w @TomWollschlager @guennemann 🧵 (1/6)
arxiv.org
Expressivity theory, characterizing which graphs a GNN can distinguish, has become the predominant framework for analyzing GNNs, with new models striving for higher expressivity. However, we argue...
1
10
19
🚀 𝗠𝗮𝘀𝘀𝗶𝘃𝗲 𝘂𝗽𝗴𝗿𝗮𝗱𝗲 𝗳𝗼𝗿 𝗤𝘄𝗲𝗻-𝗜𝗺𝗮𝗴𝗲! We've just dropped new speed, memory, and feature improvements on @replicate that are going to change your AI image generation workflow: • 𝗦𝗽𝗲𝗲𝗱 𝗯𝗼𝗼𝘀𝘁- Now generates in just 3 seconds! • 𝗠𝗲𝗺𝗼𝗿𝘆
1
5
24
🚀 Major Partnership Drop: Pruna AI × @Wiroai! 🔬 We've sped up Wiro's video generation models with our optimization technology! Our compression delivers blazing-fast inference speeds without sacrificing quality. Performance Gain: • Wan 2.2-TI2V-5B – Text-to-Video-Fast → 16
wiro.ai
A very fast and cheap PrunaAI optimized version of Wan 2.2 A14B image-to-video
0
4
15
The much anticipated image editing model from Qwen is now on Replicate https://t.co/oVMH223VtY Edit images in just 3 seconds, for $0.03 per image. We've worked with Pruna to deliver you the fastest way to use Qwen Image Edit.
12
18
203
🔥 Wan 2.2-Image is the cool new kid on the block Wan 2.2 Image generates one 2 Megapixels images for just $0.02 • 𝟮.𝟰𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗦𝗲𝗲𝗱𝗗𝗿𝗲𝗮𝗺 • 𝟭.𝟴𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗙𝗹𝘂𝘅-𝟭.𝟭 𝗣𝗿𝗼 • 𝟭.𝟭𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝘁𝗵𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆
2
2
26
We are now close to real-time video generation: 5s video generated in 5s. Try it here:
replicate.com
The fastest Wan 2.2 text-to-image and image-to-video model
@cb_doge 10 days ago, the 6 sec video took 60 secs to render, then 45 secs, then 30, now 15. We might be able to get it below 12 secs this week. This is with no reduction in visual quality. We’re working on a major upgrade to the audio track. I think real-time video rendering is
0
3
5
In the past months, we smashed all the state-of-the-art image and video models to make them x2-x10 faster! - All models from @bfl_ml: Flux-kontext, Flux-dev, Flux-schnell - All models from @vivago_ai: Hidream e1.1, Hidream l1 full, Hidream l1 fast, Hidream l1 dev - All models
1
3
8
@_akhaliq @Alibaba_Qwen @replicate Awesome! That @replicate endpoint was optimized by @PrunaAI to run super fast and be more accessible, hope you and the community likes it :)
1
1
7
love seeing more people use our optimized models
Qwen-Image @Alibaba_Qwen, a 20B MMDiT model for text-to-image generation is now available in anycoder using @replicate for generating images for your apps You can now generate images directly inside anycoder for your apps when vibe coding
0
2
2
Qwen-Image is on Replicate. With thanks to Pruna it's just $0.025 per image. https://t.co/OK9qfD1DMZ
8
12
195
Qwen-Image 3x faster on @replicate. The new Qween of open source is here and we can't wait to see what you build with it 🤩. More speed ups to come 🔜
Qwen-Image is on Replicate. With thanks to Pruna it's just $0.025 per image. https://t.co/OK9qfD1DMZ
0
6
18