Output scoring
Last updated: October 12, 2026
Output scoring judges every asset against the standards your studio writes down, and gives it a score from 0 to 100 with the reason. Use it to keep a team, a vendor or a pipeline on brief at volume.
How it works
A workspace admin writes scoring rules: each one describes a standard in plain words, the way you would brief a new artist, and can carry up to ten reference images that show the standard met. When an asset is scored, a vision model reads the rule and its reference images, looks at the asset, and returns a score from 0 to 100 with a one or two sentence rationale. Every rule in force scores separately, so an asset can be on style and still fail an IP check.
One rule is the headline score. Its number is stamped on the asset, shown in the corner of its tile, and is what the library sorts and filters by. The other rules' scores show when you open the asset.
Images, videos and 3D models can be scored. A 3D model or a layered PSD is judged from its rendered preview. Audio and text are not scored.
Score bands
Scores fall into three bands. The tile chip and the score in the asset's details take the band's colour, and the library's Score filter uses the same names.
Good
80 and above, shown in green. The asset meets the standard and can be used as it is.
Middling
50 to 79, shown in amber. Close, with something to fix: read the rationale for what.
Poor
Below 50, shown in red. The asset misses the standard and needs another pass or a different direction.
An asset the headline rule has not judged yet has no chip. On a session board, the Score filter also offers Unscored to find those.
Where rules live
Workspace rules
Set under Workspace Settings > AI Controls > Output Scoring. They apply to every project and session in the workspace. Only roles that can manage workspace settings, such as Admin, can change them.
Project rules
Set in a project's settings, under Content policy > Output scoring. A project's rules are added on top of the workspace's, which it lists as Inherited from the workspace. A project can add checks but never switch off or loosen the workspace's. A project can name its own headline rule; until it does, it uses the workspace's. Roles with the Edit content policy permission, such as Admin, Editor and Art Director, can author them.
An asset that belongs to exactly one project is scored against the workspace's rules and that project's. An asset in several projects, or in none, is scored against the workspace's rules only.
When scoring runs
Always
A rule set to Always scores every new image, video and 3D model as its generation finishes. The asset lands first and its score follows a few seconds later. Uploads are not scored automatically.
On demand
A rule set to On demand scores only when someone asks: Score now or Re-score on an open asset, a library re-score, or a request through the API or an AI assistant. Use it for checks you run before hand-off rather than on every draft.
Re-scoring
Editing a rule leaves existing scores as they were: each one records what was true when the asset was judged. Re-score the library judges existing assets against the rules as they stand now. Changing which rule is the headline does not restamp existing assets either; re-score to bring them in line.
What it costs
Each judgement is one model call per asset per rule, billed in Creative Units on what the model reads and writes, like a text generation. A rule with reference images costs more per call than one without. The library re-score asks you to confirm how many calls a run will make before it starts. See Creative Units for how usage is charged.
What scoring does not do
A score is a verdict, not a block. A poor score never stops a generation, hides an asset or prevents a download. To act on scores automatically:
- Put a Smart Image Judge (or the video, 3D or text judge) in a workflow: only outputs at or above its threshold continue, and the run fails if none pass. See Ship store art through an automated quality gate.
- Read scores through the API, the SDK or an AI assistant and route assets in your own pipeline.
- Stop a prompt or reference image before it generates with input guardrails, which are set beside scoring under AI Controls. See Your prompt or image was blocked.
What it means for you
- Write rules about what you can see in the asset. The judge reads pixels, not the prompt or the file name, so "matches the turnaround" scores better than "uses the approved prompt".
- Keep one standard per rule. A rule that checks style, model accuracy and logos at once gives one number that hides which part failed.
- Treat the score as triage. Good assets go forward, middling ones get a targeted fix from the rationale, poor ones get a new pass. An art director still signs off.