Three CycleGANs were trained to translate between domains: apple ↔ orange, horse ↔ zebra, summer ↔ winter. This embeds 300 images from each of the six domains with CLIP and asks a question those models cannot answer about themselves: how far apart are the two sides to begin with?
Cosine distance between the two domain centroids, and how often an image is closer to its own domain's centroid than to the other one. Both are computed on the raw 512-dimensional CLIP vectors, not on the 2D projection below — a projection distorts distance, so measuring on it would be measuring the projection.
Colour is the dataset; a filled dot is the first domain and a ring is the second. Hover any point to see the image. The layout comes from t-SNE over the first 50 principal components, which preserves neighbourhoods and not distances — read clusters here, read the numbers above for how far apart anything is.
Three pairs cannot establish a relationship. It is tempting to line the separability column up against how well each CycleGAN scored and draw a conclusion; with three points that would be a story, not a result. What is measured here is the domain gap itself. Whether a smaller gap makes translation harder or easier needs many more pairs than three.
CLIP ViT-B/32, 300 images per domain, seed 0, CPU only. Code and the full vectors · the models themselves