Gate results with a Smart Judge in a workflow
Last updated: October 13, 2026
Drop weak results before an expensive step: a judge scores each item against your rubric, and only the ones that clear the bar continue.
Before you start
There is one judge per kind of result: Smart Text Judge, Smart Image Judge (images and layered PSDs), Smart Video Judge and Smart Mesh Judge. Pick the one that matches what feeds it; they are set up the same way.
Judging spends Creative Units on every item it reads, and the run's estimate includes it. See Creative Units.
Steps
Open your workflow on the canvas, open the Judges group in the bottom toolbar and drag Smart Image Judge (or the judge for your media) onto the canvas.
Wire the list you want to gate into its Images input. The whole list goes in at once: a judge does not fan out, it reduces.
Write the Criteria: what a passing item looks like, in your own words. It is a standard to score against, not a prompt to generate from.
Set the Threshold, the minimum score from 0 to 100 an item needs to pass. The default is 70.
Select the node and press Run in its toolbar to try the gate on its own before you run the whole workflow.
Only the items at or above the threshold come out. Wire the judge's output into the next step so it runs on those alone. If no item passes, the workflow fails at the judge instead of running the next step on an empty list.
Result
The judge's output holds the items that passed, unchanged and in the order they came in: a PSD keeps its layers and a mesh its geometry. The node previews the first passing item and counts the rest. The judge writes no score onto the files and saves nothing to the library; the passing items reach the library through whatever your workflow saves.
Instructions that stand above the criteria
Every judge has a second text field, System Prompt, under Advanced. It holds high-level instructions on how to apply the criteria, and it is static: you write it on the node, and nothing can wire into it. When the judge reads both, the instructions rank above the criteria, so a criterion that contradicts them loses. They rank below the judge's own scoring rules, so they cannot change the 0 to 100 scale or what counts as an answer.
Use it when the criteria are not yours to write. Wire Criteria from a workflow input, so the person running the workflow describes what they want in their own words, and put the standards that must hold whatever they write in System Prompt: "Judge composition, palette and finish only; ignore the subject. Any legible text or watermark fails the image, whatever the criteria say." The rubric changes with every run, and your rules do not.
Production tips
Put a gate in front of a step that costs more than the judging: generate images, judge them, then animate or turn into 3D only the ones that pass. A gate after the last node, or between two cheap steps, adds cost and drops nothing worth the price.
Write criteria the way you would brief a reviewer: name what passes and what fails ("full body in frame, facing the camera, no text or watermark, painted texture rather than photo"). The rubrics in Scoring rule examples work here as they are.
Find the threshold before you commit to it. Score a batch you have already judged by eye with a workspace scoring rule that carries the same rubric, see where the images you consider good land, and set the threshold at the bottom of that band.
Try the gate on its own first. Run the judge node by itself on a batch you know, and read what it kept, before you run the workflow end to end.
Leave Base Model on its default unless you want a specific judge. A model you pick must be a text model that accepts image or video input.
Caveats and pitfalls
The criteria are a standard, not a prompt. A rubric written like a generation prompt ("a warrior in a snowy forest") scores how well each item matches that description, which is rarely the bar you meant.
The mesh judge sees one rendered view of each mesh. Criteria about hidden sides, topology or polygon count cannot be checked; the judge scores what it can see.
The image judge sees a layered PSD through its flattened preview. It cannot judge individual layers, layer names or layer order; the PSD passes through with every layer intact.
The video judge watches each clip whole, so a long clip takes longer and costs more to judge, and a clip too long for the judge to read fails the node rather than being skipped.
The text judge reads texts up to a few pages long. A longer text fails the node before any judging starts, so split long documents first.
The run's estimate assumes every item passes. It is the ceiling for the steps after the judge, not what a run will cost once the gate has dropped items.
An item the judge could not look at fails the node, rather than passing unjudged or being dropped unseen. A batch with a deleted file or an unsupported type stops at the judge until you fix it.
If something goes wrong
The run stops at the judge: "None of the images met the threshold". Nothing is broken. The gate rejected the whole batch rather than hand the next step an empty list, and the message names the best score it saw. Improve the generator's prompt so its results match the rubric, relax the criteria, lower the threshold, or run again for a fresh batch.
"The judge could not assess" some items. An item could not be looked at: the file was deleted, its type is one the judge cannot read, its preview had not rendered in time, or a video was too long or still encoding. Remove or replace the item; when the reason was time or encoding, run again.
"Describe what a passing item looks like." The criteria are empty. Write them, or wire a text into Criteria.
"Judging the items took too long." The batch was too large or too heavy for one run. Split the list, or judge fewer items at a time.