Skip to main content
Every image and video you generate with a character attached comes back with a continuity score out of 100. It measures how closely the take matches the character you cast in it. The score measures identity, not quality. A sharp picture of the wrong face scores low, and a dull picture of the right face scores high. Scoring costs nothing. It runs as its own job after the render settles, whichever renderer produced the take, and the wait includes starting a GPU. The card shows Queued and then Running while it waits, and stops waiting if no score has arrived after four minutes.

The bands

The bands are provisional. The thresholds were set on local output and have not been calibrated against hosted renderers, which is why the card says so on every take. Compare two takes of the same character against each other rather than reading one number on its own.

What the number is made of

Two signals go into the score:
  • Face measures the face itself, using SFace.
  • Subject measures the whole person, including build and pose, using DINOv2.
Wardrobe is measured and reported beside the score, never inside it. Clothes change between shots on purpose, so letting an outfit change pull the identity score down would make the number useless. A take that measured only one of the two says so, because a partial measurement is not the same claim as a full one: The card also says whether it scored against the character’s original references or its current ones, which differ once you have rebuilt it.

Video

A clip is sampled frame by frame, and the headline is the median across the frames where a face was found. Beside it you get the lowest frame and the second it happened, because a median hides a single visible break. Click a sample to jump the player to it. Frames where no face was found are counted and reported separately, never scored as zero. A turned head is not a wrong face.

More than one character

Cast several characters in one run and each is scored separately, with the frames where that character went missing and the frames where two characters swapped reported alongside. One average across the whole cast would hide a swap.

What is not scored

Audio and trained LoRA files have no likeness to measure, so they are not scored. Neither are runs with no character attached: there is nothing for the take to have matched, so no card appears.

Next steps

Finetuning

Fix the references behind a low score.

Pricing

What a generation costs. Scoring is free.