The eigencrungi
Move the sliders to combine patterns of variation across 24,400 archive images. The display starts with their mean image. Select an archive thumbnail to compare its reconstruction with the original. Images are centre-cropped and resized to 128 × 128 pixels before analysis.
Numbered sliders are principal components fitted to the image pixels. Their stability scores describe how closely they reproduce across subsets of the archive. The separate “crungus-ness” slider follows similarity to CLIP's text embedding of “crungus”, with a pixel visualisation fitted from the archive.
These sliders explore variation in CLIP image features, which can reflect subjects and visual style. Each filmstrip shows Kandinsky 2.2's interpretations of positions along one direction. Compare those generated previews with the nearest archive images below the main display.
The preview follows the last slider you moved; it does not combine sliders. The nearest archive images respond to all slider positions together. Kandinsky is also one of the 122 image models in this collection.
loading the eigencrungi…
This experiment adapts the eigenfaces approach introduced by Turk and Pentland in 1991: represent images as a mean plus weighted principal components. Here, subjects are not aligned before analysis, so components capture differences in brightness and framing as well as depicted subjects. Retaining a limited number of components produces an approximation of each image.
Stability is the cosine similarity between corresponding components fitted to disjoint halves of the archive, averaged over eight splits. Higher scores indicate closer agreement across those splits. We name components only when their score exceeds our chosen threshold of 0.7. Scores below that threshold indicate less reproducible directions, which remain available to explore. Read about the method or see the historical experiments.
Each image is represented by 1,280 features from open-clip's ViT-bigG-14 model. Principal component analysis finds directions of variation in these features. Stability is measured across eight disjoint-half splits, using the same procedure as the pixel analysis.
Filmstrips sample a single direction at seven positions between −3 and +3 standard deviations. At each position, Kandinsky renders three candidates; we retain the middle candidate ranked by embedding similarity to the target. These are decoder interpretations of positions in feature space, including positions no archive image occupies. The adjacent thumbnails show collected outputs for comparison.
We render filmstrips for components whose stability exceeds 0.7. Other sliders still affect the nearest-image results. The text-similarity slider uses CLIP's association with “crungus”; it is distinct from the consistency scores on model pages. Read about the method or see the decoder configuration notes.