2026.09 · Pebblous Data Communication Team

Reading time: ~5 min · Explorer interface in Korean · 한국어

Summary

Give the same 300 leaf photos to four AI models and each returns a different list of "similar" photos. Rather than describe that difference, this article opens the explorer so you can press on it yourself.

The explorer holds 300 photos of common-bean leaves taken with smartphones in Ugandan fields: 100 each of two diseases (angular leaf spot and bean rust) and healthy leaves. Four AI models turned each photo into a high-dimensional vector; those vectors are laid out on a 3D map. Pick a point and the eight photos that model considers closest appear at full resolution. Switch the lens and the neighbors change.

For three leaves, the explorer also stores the actual intermediate values at 30 stages inside the network, from input tokens through 24 transformer blocks to the final normalization, so you can step through where the three start to diverge. For six leaves you can open a patch-by-patch similarity map across 196 patches.

Every photo, coordinate, neighbor, and layer value on screen is a stored measurement. Examples built with mathematics to explain a principle are labeled synthetic. The explorer is not a crop-disease diagnosis service, and a similarity map is not an explanation of why the model sees things the way it does. The interface is in Korean; the data files and this article are enough to follow along.

Explorer

Use it right here, or open it in a new window on a narrow screen. About 100 MB of data is downloaded once.

Open in a new window ↗

No model runs inside the explorer; everything shown was computed and stored in advance. Drag to rotate the 3D map; click a point to change the neighbors.

How to use it

Switch the lens

In the embedding section (임베딩 탐색), change the model. The whole arrangement of 300 points shifts. The first thing you notice is how differently the three colors, two diseases and healthy, separate under each lens.

Pick a point and look at its neighbors

Click a point to see the original photo and the eight photos that model ranks closest. Are they grouped by the shape of the lesions, or by leaf color and background? Click the same point under another lens and the answer changes.

Step through the layers

In the layer section (층별 텐서), follow one leaf of each class through 30 stages. Watch which block widens the distance between the three, using the stored numbers themselves.

Open one photo as 196 patches

In the patch section (실제 패치), choose one of six photos and click a patch. The other 195 patches are colored by how close their representations are to the one you chose. That is a distance between patch representations, not a record of where the model looked.

What you can see

1. Same leaf, different neighbors

Each model has its own notion of "similar." Photos grouped by lesion shape under one lens are grouped by leaf color and background under another. Watching eight neighbors swap out as you switch lenses makes one point concrete: an embedding is not the meaning of a photo but the photo as one particular model sees it.

2. Inside the layers, not only at the output

The final vector only tells you how far apart three leaves end up. Stepping through the 30 stored stages shows at which block the difference grows. The point where representations diverge is not the moment the model "recognized the disease"; where a representation changes and where a decision is made are different questions.

3. Does a better lens make a better selection?

Pebblous has already run a selection experiment on these 300 photos, choosing 120 for training. In the space of the large DINOv3 model, the 120 selected photos gave 3.35 percentage points higher balanced accuracy than 120 photos drawn at random with matched class counts. That figure holds for that lens and that evaluation setup; it does not generalize to other lenses or other data. The explorer's results section lists the conditions and the uncertainty.

What it cannot tell you

  • It is not a diagnosis. The explorer does not judge disease from a leaf photo. Labels were assigned by experts at collection time, and no pathogen was confirmed from photos alone.
  • Similarity is not a reason. Patch maps and neighbor lists are distances between representations, not attention and not a causal explanation.
  • The 3D map shows only part of the structure. A 1,024-dimensional vector is squeezed onto three axes. The eight neighbors are computed in the original space, not on the map.
  • 300 photos are not the whole dataset. They are a subset of the 1,295-image public dataset, 100 per class from the training split. Layer values exist for three photos and patch maps for six.
  • Synthetic examples are not measurements. The EMA trajectory and Gram experiments that illustrate the training principle are mathematical models, labeled synthetic on screen, not training logs.

Data, sources, rights

Photos
300 images from the training split of the iBean dataset (100 per class, 500×500), collected with smartphones in Ugandan fields by Makerere AI Lab and the National Crops Resources Research Institute (NaCRRI) and labeled by experts on site. The source repository is MIT licensed. iBean repository · TensorFlow Datasets catalog
Models
The four lenses are DINOv2-L, DINOv3-L, SigLIP 2, and C-RADIOv4-H. The reference lens is Meta's large DINOv3 model (ViT-L, 1,024 dimensions); layer values and patch maps were stored from it. No model weights ship with the explorer and no model runs in it. DINOv3 paper (2025) · Official code and model card
Computation
The 3D map is a PCA projection of L2-normalized, mean-centered vectors onto three axes. Neighbors are the top eight by cosine similarity, computed exhaustively in the original space. Procedures and file hashes are in the explorer's methods section.
Downloads
Research report (Markdown, Korean) · Embeddings and neighbors (JSON) · Evidence ledger (JSON). Reuse of the photos and model-derived values follows the licenses of their sources.
How it was made
The explorer is a snapshot, dated September 25, 2026, of a research workbench Pebblous built while running data-selection experiments. Code and copy were written with the help of AI coding tools; data verification and the decision to publish were made by people.

Tangible Data Story — articles built on data you can handle

This is the first Tangible Data Story, the data-journalism strand of Tangible Data. The other strand, Tangible Data Art, lives alongside it in the hub.

Go to the Tangible Data hub →