Jannik retweeted
⚡️Few-step diffusion without a teacher, self-distillation or JVPs.
iDDM: a single model trained from scratch in one stage. ImageNet-256: 4.48 FID at 4 steps and 2.38 at 50, from the same checkpoint
🧵 (1/6)
Jannik retweeted
Real-world scenes contain many objects moving simultaneously.
Yet current video representations entangle motion globally, exactly what we don’t want for object-centric motion understanding and generation.
We introduce localized context-aware motion representations.🧵👇
Jannik retweeted
Join us for our short oral today at 10:12 (in Palissad) at ECCV, and our poster (#442) in poster session 3 (morning session)
Directly predict the distribution over possible futures in one forward pass.
To find out how, come by our poster at #ECCV in #Malmö today.
Poster session 3 - Poster 101. See you at 10:30 am!
Jannik retweeted
An academic lab is nothing without its researchers.
Thank you to all the students and postdocs who welcomed me into your labs and shared your time, ideas, and enthusiasm. Working, learning, sharing meals, and playing 🏓 alongside you was a major highlight of my sabbatical.
That’s a wrap on my sabbatical. Huge thanks to my wonderful hosts, @KaterinaFragiad (CMU), Björn Ommer (LMU), and @davsca1 (UZH). World-class researchers, and even better people!
Thank you for an unforgettable year ❤️
TL;DR
We shouldn’t think of the future as a single trajectory. Following Schrödinger's famous analogy 🐈: We should think of it as the space of possible outcomes. Once you have that representation, planning and uncertainty estimation become different queries on the same object.
1
6
75
Joint work with Timy Phan, supervised by Björn Ommer.
Paper: arxiv.org/abs/2607.25984
Project Page: compvis.github.io/schroeding…
Code: github.com/CompVis/schroedin…
See you at #ECCV 2026!
Jannik retweeted
Most representation learning methods produce generic embeddings that try to preserve everything in an image.
We introduce VICIS: a way to use example sets to define a tailored embedding space for what matters.
This is useful whenever the desired visual signal is easier to show than to describe.
Code & weights available now 🧵👇
How can we build uncertainty-aware compression stages for latent space forecasting?
Check out Poster #401 today at 4:45pm at #CVPR !
Jannik retweeted
Probabilistic Precipitation Nowcasting with Rectified Flow Transformers
🕐Sat 6th, 4:45 - 6:45 PM
📍ExHall A & F 401
teapawt.net/jo_schb/status/2062333…
Check out our new CVPR paper on uncertainty-aware first-stages for precipitation nowcasting! ⛈️
⚠️ Standard first stages are not sufficient for safety-critical applications!
The most extreme weather events are often the hardest to decode.
One latent → many plausible reconstructions
Deterministic decoders hide that uncertainty.
Meet FREUD 🧵👇
1
41
Jannik retweeted
The internet is full of video. So why can't novel view synthesis just scale on it?
Real-world video is simultaneously unposed, messy, and dynamic, breaking self-supervised NVS.
We fixed that. RayDer learns static-scene NVS from dynamic internet video, scaling like an LLM. A🧵
