Walkthrough · one measurement

Finding a candidate source among 1.28 million.

One worked example, on images, of what generation-moment attribution returns and how to read it. Four steps and a note on what it does not claim.

STEP 1

The haystack

DMD2, a state-of-the-art diffusion model, was trained on 1.28 million photographs from ImageNet. This is a public corpus and a public model, which is why we can show the whole exercise rather than describe it.

The question a licence needs answered is which of those 1.28 million works contributed to any particular new image. Asked after the fact, against a finished artifact, it is a needle-in-a-haystack problem with no ground truth to check against.

A grid of ImageNet training photographs 1,280,000 images
STEP 2

The generated item

We asked the model for new images of ostriches and picked one to investigate. The image is new: it is not in ImageNet, and it is not a copy of anything in it.

This is the ordinary case, and the hard one. Nothing here would be caught by a duplicate check or a watermark, and nothing about the image announces which works shaped it.

An image generated by the model
Generated by DMD2. Not present in the training corpus.
STEP 3

The candidate source returned

Out of the 1.28 million training photographs, Yodal returns the photograph on the right as a candidate source for that generation, in under a second. We call it a candidate source rather than the source deliberately: what is reported is measured contribution with a stated basis, not a causal proof.

Generated Generated image
Candidate source Candidate source photograph from the training corpus
STEP 4

Why this is not chance, and not a resemblance match

The two images are not identical, and the candidate was not chosen for looking like the output. What makes the result meaningful is that it coincides with the generated image on four independent properties, each of which most of the corpus does not share.

Same subject
Both are ostriches. Roughly one in a thousand ImageNet photographs is, so drawing one at random gives you this much and no more.
A random sample of ImageNet photographs
Same pose
Both birds are lying down. Most of the thousand candidates of this subject are standing.
Standing ostriches from the corpus
Same background
The settings closely correspond, where most candidates of the same subject and pose differ.
Ostriches with varied backgrounds
Same framing
The crop and camera distance match. Many candidates show only a head or a neck.
Ostriches with varied framing
Reading the result honestly

What this shows, and what it does not

It shows that contribution can be measured at the moment of generation, at corpus scale, fast enough to sit inside a transaction. That is the claim, and it is the one a licensing regime needs.

It does not show a causal proof that this photograph and no other produced this image.

It also does not settle whether this output infringes. The regime is designed so that a rightsholder's terms decide what happens next rather than our threshold deciding it for them.