Paimon 0.16Real frame · real model output
That's Paimon.
The model is 16% sure.
You knew instantly. A computer does not — and teaching one costs thousands of examples that a person has labelled by hand. crowdmon makes those examples cheap: it turns recorded gameplay into frames, has a model guess at every one, and cuts the human job down to yes, no, or nudge.
How it works
Three steps, and only one needs a person
From a link you paste to a dataset somebody could train on.
Cut the video up
A worker pulls the video, slices ~2,700 stills, throws away the near-duplicates and keeps a random 200 spread across the timeline.
Take a rough guess
An open-vocabulary detector proposes a box for anything it takes for a character. It has never been shown these characters — it works from their names.
Rule on it
Accept, adjust, or reject. Seconds per frame instead of minutes. Verdicts are append-only and packaged into a snapshot with a train/test split.
Why it works
Checking is cheap. Drawing is not.
The 2023 version failed here. It had a perfectly good annotation tool where every box was drawn from scratch, so hour ten cost exactly what hour one cost. There is no finish line in a job shaped like that.
A bad guess is still useful. Rejecting one takes under a second, and a rejection is a label too. That is why the detector being weak is not a problem to be fixed before the rest can work.
So the model is the swappable part. It sits behind a one-method interface and can be replaced in a single file. The pipeline, the queue and the export around it are the thing that was built.
FAQ
Questions worth asking
Is the detector any good?
No, and that is deliberate. It is a zero-shot bootstrap that has never seen these characters — confidences sit around 0.10–0.20 and plenty of its boxes are wrong, including one on this page. Accuracy is not the deliverable; throughput per human minute is.
Do verdicts on the demo change the dataset?
Not the anonymous demo's. Those verdicts are recorded and tagged, then excluded when a snapshot is built — trying it out never enters the dataset. Signing in to contribute is different: your rulings are recorded under your own account, and count toward the dataset once an admin has trusted it. Admitting untrusted labels by default would force consensus resolution, agreement scoring and trust weighting — all deliberately out of scope.
Can it find something the model missed entirely?
Not by itself. A frame with no prediction never enters the queue, so a miss looks identical to an absence. Admins can file missing-object reports, and the report rate per class is the number that says whether a prompt is working.
Does it train anything?
Not yet. Training, a model registry and a distilled detector are all out of scope for now. What exists is the part that produces the data those would need.
Why Genshin Impact?
Recognisable, consistently-rendered subjects and an endless supply of source video, which makes it a good harness for the pipeline. The frames are game screenshots and the dataset is not distributed as a product.
See it get one wrong
One real frame, the detector's real output, three buttons. About ten seconds, no account.