Our dtour is accepted to IEEE VIS ’26 and win a Best Short Paper Honorable Mention 🎉

Our paper on interactively touring through many 2D projections of high-dimensional data was accepted to IEEE VIS ‘26 and received a Best Short Paper Honorable Mention! 🎉

Preprint: https://arxiv.org/abs/2605.04306 Try dtour: https://dtour.dev Code: https://github.com/flekschas/dtour

But who still reads papers when they could spend their precious time understanding high-dimensional data? To celebrate the paper and make the second part easier, a new version (0.5) of dtour is out, which lets you render up to 32 previews, link multiple views via selection, and have AI build tours through MCP.

Increasing previews to 32 (and uniformly subsampling for 32+ keyframes) makes it easier to study long sequences. For instance, @Mingxuan Zhang and colleagues introduced SHERLOCK, an interesting method that learns structured representations of single-cell perturbation data. I was curious how the learning progresses, but 16 previews were too limiting. With 32, you can now see more clearly when structure emerges while training. (In the 2D UMAP, most visible structure forms by about epoch 250 and then barely changes. But a single 2D layout can only show so much, so I also toured two 4D UMAPs. See the video!)

Linking previously worked only in notebooks. Now you can do it easily on the webapp as well. For instance, go to https://dtour.dev/?dataset=digits to see a PCA tour of hand-written digits linked to a classic 2D UMAP. Try selecting some points!

But how do you create such tours? To make it even easier to get started, dtour now comes with an MCP server. For Claude Desktop, it’s a one-click install: download https://github.com/flekschas/dtour/releases/latest/download/dtour.mcpb and double click it. With that in place you can re-create the linked digits by asking Claude to “use dtour to show a PCA tour of the pixel columns in https://data.dtour.dev/digits.csv next to a 2D UMAP, colored by digit.”

Overall, Nezar Abdennur and I built dtour to make it as accessible and effortless as possible to visually traverse high-dimensional space. There are many useful 2D embedding methods, but no single 2D view can capture every angle of a complex manifold. Being able to easily move through many projections can give better intuition, and MCP gets you started in seconds.

If dtour proves useful in any dimension of your reality, please share your favorite view and have a great day.

Also super excited to catch up with many friends at VIS next month! 👋