Lineage

Lineage tells you whether an image was made with Juggernaut XL v9 or Stable Diffusion XL 1.0, or says it can't tell. It gives an answer for about 9 images in 10, and when it answers, it is right more than 9 times in 10.

It knows only these two models: an image from any other model may still be named as one of them, or get "cannot distinguish"; neither means it came from these models.

Measured by re-scoring, through this service, the 3,718 images of a sealed, independently audited experiment. Original 1024×1024 PNG images, generated with 25 steps and guidance 7.5 from prompts that name no real person.

Request access

The measured statement, as GET /v1/contract serves it

per_true_model

juggernaut-xl-v9

answered: 1750
correct: 1608
correct_given_answered: 0.9188571428571428
correct_given_answered_wilson95: [0.9051274033496538, 0.9307519599862367]
coverage: 0.9413663259817105
coverage_wilson95: [0.9297484405227134, 0.9511638173955121]
images: 1859

what

A compatibility check: the confirmation's own images, re-scored through this service under key-policy-1 and judged by the confirmation's test. It is not an independent confirmation: the images are the ones the confirmation scored.

applies

Only to a job whose scope declaration is affirmed: original, unmodified 1024x1024 PNG outputs of the two checkpoints, at 25 steps and guidance 7.5, from prompts that name no real person. Nothing was measured about any other image.

reading

Each rate is for the images one model made. Coverage is the share of them the gate answered. Correct given answered is the share of those answers that named that model.

per_true_model

sdxl-1.0-base

answered: 1684
correct: 1571
correct_given_answered: 0.9328978622327792
correct_given_answered_wilson95: [0.9199355141390102, 0.9438896220586046]
coverage: 0.9058633674018289
coverage_wilson95: [0.8917389217479732, 0.918313849244472]
images: 1859

what

applies

reading

Each rate is for the images one model made. Coverage is the share of them the gate answered. Correct given answered is the share of those answers that named that model.

contract: contract-3
images: 3718
result: PROMISE_KEPT: PROMISE_KEPT in both directions

instrument

the instrument this was measured on, and the commit whose scoring code it ran. It is a statement about results made on that instrument by that code, and on no other

instrument_id: 1970970acf6ef103
runtime: pod
code_commit: 5ad06109639d776b0d992e03489eaf0f78476e51

source

verdict_file: runs/pod_session/2026-10-03/mvp-1-compat/session-b/verdict.json
sha256: dffae185fa438fbf2c12a3dda5c552b3e29fb264843930d3e014f9981c4aa7f8
commit: 21d99f225bd24c9341aae6b83ba5784a8c32d38b

The four outcomes, as the contract names them

juggernaut-xl-v9

the frozen pair gate answered, and it named Juggernaut XL v9 over Stable Diffusion XL 1.0 base

sdxl-1.0-base

the frozen pair gate answered, and it named Stable Diffusion XL 1.0 base over Juggernaut XL v9

cannot_distinguish

the frozen pair gate abstained. It does not mean neither model, an unknown generator, or equal probabilities

unsupported_input

an observable input requirement or the required scope declaration failed, and nothing was scored. It is not unknown-model detection

The samples, as screened

The selection rule

the first 12 images, in ascending order of the file's sha256, that pass the screen; nothing is chosen.

The skip rule

Skip any image showing a recognizable character, brand or logo, a real person, a weapon or a disaster, and take the next.

5 were skipped on the way, each for what it shows and none for how it scored: a disaster (1), a face that could be taken for a real person (1), a recognizable character (2), a weapon (1).

The counts shown, for the 12 on the page

named correctly 9, named wrongly 1, no answer 2.