Every reference FIELD pinned for a client, hanging below the centre of what was collected, at its own measured distance from it. Every image FIELD published for that client, hanging below the research that best explains it. The distance between the two is the part of the work the references do not account for — and half of it, measured against forty unrelated clients, turns out not to be about that client at all.
Both reproduce the same sheet. The token carries every parameter that fixes the render; --state decodes it, so a composition found here comes back byte for byte on paper.
loading the field…
One lightly drawn territory per client. Every vertical length is a cosine distance in 512-dimensional CLIP space. Horizontal position is local density and nothing else — the sideways distance between two marks measures nothing at all.
The line across the top is the centre of everything the studio collected for that client. A reference's depth below it is that reference's own distance from that centre, so the depth of the cloud is the real spread of the research — and it is wider than the gap the piece is about. The distance from research to output is smaller than the distance across the research itself.
Each drops from a reference neighbourhood to the output neighbourhood it best explains, and the drop is that pair's own measured gap. Thickness is how many individual references actually point at that output group; a broken line means either none of them do, or the group is too small to resample and the number cannot be checked.
The solid rule is 1 − pull, the client-level gap. The dashed rule is what the same references give against an unrelated client's published work. Where the dashed rule sits above the solid one, the relationship is doing no work, and the territory is drawn as a ghost.
A reference drawn as a ring rather than a dot sits closer to at least half of the other thirty-nine clients' published work than to its own client's. 455 of 1,825 do — none at all at Adidas, 43% at Google, 93% at Fairlight. It is the same story the permutation test tells, derived independently, which is the reason to believe either.
Click any mark to pin it. The cosine printed between the two slots is recomputed in the browser from the same 512 dimensions the sheet is drawn from — not a lookup into a table of answers, so you can check any pair the piece implies.
A high pull is not evidence on its own. Two sets of photographs have centroids close together whatever the relationship between them. So every client's own number is tested against 40 client output centroids — every client on field.io with at least three embedded published images, including the twenty-eight whose reference side never joins.
Nothing here survives a Bonferroni correction. Twelve tests want p ≤ 0.0042 and an exact permutation over forty clients cannot go below 0.025. Four hits at p ≤ 0.05 where 0.6 are expected by chance is a real collective signal and it is not a per-client licence. The four indistinguishable territories and the two inverted ones are drawn as ghosts for that reason, and they are not hidden.
The bootstrap widths are not comparable. They run from 0.021 for Adidas to 0.184 for VASHI, and only 25 of the 66 client pairs on this sheet have 95% intervals that separate at all. Most of the ordering you can see is not a distinction the data supports.
Titles match on 2 of 166 channels, so the join is by client name. A client with several jobs has both sides pooled. Nothing on this page connects a reference to the project it was pinned for, because that join does not exist.
It is filtered by taste and by client approval. The lower cloud is what was shown, never what was made, so a wide gap can be a publishing decision rather than a design one.
pull tracks output count at Spearman +0.61. Cutting every client to the smallest one's six images shifts the level by up to 0.036 and leaves the order nearly intact (+0.91), so the ranking survives the confound and the absolute values do not.
The model has its own reasons for calling two pictures close. Every claim here is a claim about one embedding space, and it is worth looking at rather than trusting.
python3 src/viz/translation_field.py --list
python3 src/viz/translation_field.py --page a2l --dpi 300
python3 src/viz/translation_field.py --client Adidas --page a3l