Run Inference

Run models on images, video, and datasets from the terminal

eyepop run is the core workflow. A flag names the target (what to run) and --media-path names the media (what to run it on):

eyepop run --model <alias-or-uuid> --media-path <media>

Targets

Target	Meaning	Example
--model <alias-or-uuid>	A published model or ability	--model eyepop.person:latest
--pop <pop>	A Pop from eyepop get pops, an ability, or the Pop itself as JSON/YAML (inline or a file)	--pop people-common-objects, --pop ./pop.json
--session <uuid>	An existing session or deployment	--session 73a991ac-...

Exactly one target per run. A published model is the usual choice — list what you can run with eyepop get models, and published abilities with eyepop get abilities.

On a machine running an on-premise instance, you can omit the target entirely and the run uses the pop that instance serves.

Prefer a guided flow? eyepop tui lists your runnable models and walks you through picking media and running. Running eyepop run with no target and no media prints its help.

Run on media

Run a model on a single image:

eyepop run --model eyepop.person:latest --media-path image.jpg

Ask a VLM ability a question about the image:

eyepop run --model <vlm-ability> --media-path image.jpg --prompt "Describe this image"

Run several sources, or a directory recursively:

eyepop run --model eyepop.person:latest --media-path image1.jpg --media-path image2.jpg
eyepop run --model eyepop.person:latest --media-path ./images --recursive

--concurrency runs several inputs at once. --no-cache forces fresh inference instead of reusing a cached result; it applies to VLM ability runs, with or without --prompt, and eyepop run refuses it on --pop, --session, and any --model alias that resolves to a published model.

Evaluate on a dataset

Scoring against a dataset is its own command — eyepop evaluate takes an ability and a dataset:

eyepop evaluate --ability my-namespace.find-kittens:latest --dataset people

Scope to a partition, or return immediately with a request id instead of waiting:

eyepop evaluate --ability my-namespace.find-kittens:latest --dataset people --partition test
eyepop evaluate --ability my-namespace.find-kittens:latest --dataset people --no-wait

Check or watch the evaluation:

eyepop get evals <request_id>
eyepop get evals <request_id> --watch

Run a Pop or an existing session

eyepop run --pop people-common-objects --media-path image.jpg
eyepop run --session <session_uuid> --media-path image.jpg

A Pop you wrote yourself runs straight from its file: eyepop run --pop ./pop.json --media-path image.jpg. To keep it warm between runs, create a deployment from it with eyepop create deployment --pop ./pop.json (see Persistent Sessions) and run against it with --session.

Pops and sessions take media only. To score against a dataset, use eyepop evaluate.

Next steps

• Command Reference — every run and evaluate flag

• Resources — find the models, abilities, and pops to run