Thomas Wimmer

DUSt3R et al. are impressive, but how do they actually work? We investigate this in our project 𝘜𝘯𝘥𝘦𝘳𝘴𝘵𝘢𝘯𝘥𝘪𝘯𝘨 𝘔𝘶𝘭𝘵𝘪-𝘝𝘪𝘦𝘸 𝘛𝘳𝘢𝘯𝘴𝘧𝘰𝘳𝘮𝘦𝘳𝘴!⁣

We share findings on the iterative nature of reconstruction, the roles of cross and self-attention, and the emergence of correspondences across the network [1/8] ⬇️

ICCV 2025 🌺 Aloha from Hawaii! MPI-INF (D2) is presenting 4 papers this year (one Highlight). Thread 👇

I am on my way to #ICCV2025 to present DIY-SC, where we refine foundational features for better semantic correspondence performance. Please come by our poster poster #538 (Session 2) if you're interested or want to chat about my latest project, AnyUp!

🤔 What if you could generate an entire image using just one continuous token?

💡 It works if we leverage a self-supervised representation!

Meet RepTok🦎: A generative model that encodes an image into a single continuous latent while keeping realism and semantics. 🧵 👇

AnyUp Universal Feature Upsampling

Try it out now! Code and model weights are public.

💻 Code: github.com/wimmerth/anyup

Great collaboration with Prune Truong, Marie-Julie Rakotosaona, Michael Oechsle, Federico Tombari, Bernt Schiele, and @janericlenssen.bsky.social!

CC: @cvml.mpi-inf.mpg.de @mpi-inf.mpg.de

Generalization: AnyUp is the first learned upsampler that can be applied out-of-the-box to other features of potentially different dimensionality.

In our experiments, we show that it matches encoder-specific upsamplers and that trends between different model sizes are preserved.

When performing linear probing for semantic segmentation or normal and depth estimation, AnyUp consistently outperforms prior upsamplers.

Importantly, the upsampled features also stay faithful to the input feature space, as we show in experiments with pre-trained DINOv2 probes.

AnyUp is a lightweight model that uses a feature-agnostic layer to obtain a canonical representation that is independent of the input dimensionality.

Together with window attention-based upsampling, a new training pipeline and consistency regularization, we achieve SOTA results.

Foundation models like DINO or CLIP are used in almost all modern computer vision applications.

However, their features are of low resolution and many applications need pixel-wise features instead.

AnyUp can upsample any features of any dimensionality to any resolution.

Super excited to introduce

✨ AnyUp: Universal Feature Upsampling 🔎

Upsample any feature - really any feature - with the same upsampler, no need for cumbersome retraining.
SOTA feature upsampling results while being feature-agnostic at inference time.

🌐 wimmerth.github.io/anyup/

Architecture for Unpaired Multimodal Learner.

Suppose you have separate datasets X, Y, Z, without known correspondences.

We do the simplest thing: just train a model (e.g., a next-token predictor) on all elements of the concatenated dataset [X,Y,Z].

You end up with a better model of dataset X than if you had trained on X alone!

6/9

Happy to find my name on the list of outstanding reviewers :]

Come and check out our poster on learning better features for semantic correspondence in Hawaii!

📍 Poster #538 (Session 2)
🗓️ Oct 21 | 3:15 – 5:00 p.m. HST

genintel.github.io/DIY-SC

What was the patch size used here?

All the links can be found here. Great collaborators!

bsky.app/profile/odue...

🚀 Just accepted to ICCV 2025!

In DIY-SC, we improve foundational features using a light-weight adapter trained with carefully filtered and refined pseudo-labels.

🔧 Drop-in alternative to plain DINOv2 features!
📦 Code + pre-trained weights available now.
🔥 Try it in your next vision project!

The CVML group at the @mpi-inf.mpg.de has been busy for CVPR. Check out our papers and come by the presentations!

Hello world, we are now on Bluesky 🦋! Follow us to receive updates on exciting research and projects from our group!

#computervision #machinelearning #research

Gaussians-to-Life: Text-Driven Animation of 3D Gaussian Splatting Scenes We introduce a method to animate given 3D scenes that uses pre-trained models to lift 2D motion into 3D. We propose a training-free, autoregressive method to generate more 3D-consi...

We only use open-sourced models and the implementation of our method is readily available. Please check out the paper website for more details:

wimmerth.github.io/gaussians2li...

We can animate arbitrary 3D scenes within 10 minutes on a RTX4090 while keeping scene appearance and geometry in tact.

Note, that since the time I worked on this, open-sourced video diffusion models have improved significantly, which will directly improve the results of this method as well.

🧵⬇️

Improvement of multi-view consistency of generated videos through latent interpolation. In addition to the rendering of the dynamic scene f, using the rendering function g from the current viewpoint g(f)_s, we compute the latent embedding of the warped video output v_{s-1} of the previous optimization step (from a different viewpoint). We linearly interpolate the latents before passing them through the video diffusion model (VDM), which is additionally conditioned on the static scene view from the current viewpoint. The resulting output is finally decoded to a new video output v_s.

While we can now transfer motion into 3D, we still have to deal with a fundamental problem: Lacking 3D consistency of generated videos.
With limited resources, we can't fine-tune or retrain a VDM to be pose-conditioned. Thus, we propose a zero-shot technique to generate more 3D-consistent videos!
🧵⬇️

Method overview for lifting 2D dynamics into 3D. Pre-trained models are shown in blue. We detect 2D point tracks and use aligned estimated depth values to lift them into 3D. The 4D (dynamic 3D) Gaussians are initialized with the static 3D scene input.

Standard practices like SDS fail for this task as VDMs provide a guidance signal that is too noisy, resulting in "exploding" scenes.

Instead, we propose to employ several pre-trained 2D models to directly lift motion from tracked points in the generated videos to 3D Gaussians.

🧵⬇️

Had the honor to present "Gaussians-to-Life" at #3DV2025 yesterday. In this work, we used video diffusion models to animate arbitrary 3D Gaussian Splatting scenes.
This work was a great collaboration with @moechsle.bsky.social, @miniemeyer.bsky.social, and Federico Tombari.

🧵⬇️

Can you do reasoning with diffusion models?

The answer is yes!

Take a look at Spatial Reasoning Models. Hats off for this amazing work!

I wonder to which degree one could artificially make real images (with GT depth) more abstract during training in order to make depth models learn these priors that we would have (like green=field, blue=sky) and whether that would actually give us any benefit, like increased robustness...

Ah, thanks, I overlooked that :)

Nice experiments! What model did you use?

🏔️⛷️ Looking back on a fantastic week full of talks, research discussions, and skiing in the Austrian mountains!

Give a warm welcome to @janericlenssen.bsky.social!

Well well, it turns out that GIFs aren't yet supported on this platform. Here is the teaser video as an MP4 instead:

This work was led by @mohammadasim98.bsky.social and is a collaboration with Christopher Wewer, Bernt Schiele and Jan Eric Lenssen.

Check out the website with lots of nice visuals that show how our metric works and use it in your next diffusion model project!

geometric-rl.mpi-inf.mpg.de/met3r/