CV models and local weights

A model-backed method is ready only when its runtime, model files and method configuration are all available. An installed ml.runtime does not mean weights are already downloaded, while a file on disk does not prove runtime compatibility.

Network boundary. Downloads start only from an explicit button and contact Ultralytics/GitHub or a pinned Hugging Face repository. The model host sees ordinary device HTTPS metadata such as IP and User-Agent, but Screph does not send it the canonical project, markup or source frames. Inference is local after download. For a sensitive network, use a manually prepared or offline runtime/model pack.

Three independent layers

Layer	Where to check	What it proves
`ml.runtime`	`Настройки → Зависимости`	Availability of the selected Python/ML runtime and its packages.
Model asset	The model family's settings page	Required files, path and, when a pin is declared, SHA-256.
Active backend	Model settings and the CV workspace	Which ready model a specific method uses.

The Check runtime and Dependencies... controls on model tabs use the same shared ml.runtime component; they do not install a hidden second Python copy.

YOLO

Open Settings → YOLO Models. The page stores the model directory, GPU flag, autoload option and active models for segment, detect and pose consumers. Its table shows name, type, size, estimated RAM/speed and actual file status.

Download selected runs Ultralytics in a background worker and moves the resulting .pt file into the selected directory.
Delete from disk removes the local file after confirmation; the registry entry remains available for another download.
Import custom model copies a user-trusted .pt and asks for its name and detect/segment/pose type.
Search, filtering and sorting do not modify files. Double-click opens help for a registry entry.

Current catalog boundary. The bundled registry contains segment and pose entries (including YOLOv8/YOLO11 and FastSAM), but no ready-made detect entries. Active detect selection and custom detect import are implemented. A custom entry lives in the current process registry; the copied .pt remains on disk, but its entry must be added/imported again after restart. An empty active-model list does not prove that a default filename is ready.

SAM

Open Settings → CV Settings → SAM Models. This page selects the directory, default backend, autoload behavior and each backend checkpoint.

Backend	Checkpoint provisioning	Check
MobileSAM	Managed download from a pinned Hugging Face revision.	Presence, SHA and a backend load test.
EfficientSAM	Managed download from a pinned Hugging Face revision.	Presence, SHA and a TorchScript load test.
FastSAM	The button shows instructions: automatic download is disabled and the checkpoint must be obtained manually from the official source.	Presence and a load test; no SHA pin is declared for this manual artifact.

OmniParser

Open Settings → CV Settings → OmniParser. This page manages the directory, GPU, PyTorch/CUDA system information and the multiple required files of each model component.

The recommended V2 set consists of icon_detect_v2 and icon_caption_v2; legacy V1.5 entries also remain in the registry.
Download selected obtains artifacts from the pinned revision of the corresponding Hugging Face repository.
Downloaded requires every mandatory local file; one missing file makes the component unready.
Each registry entry shows its license, which may differ between detector and caption components.

What SHA status means

Не загружена / missing — a file or part of a required set is missing.
Загружена / valid — required paths exist; a full hash check may not have been requested yet.
Ошибка / broken — structure or a declared SHA did not match; replace or re-download the asset.

Check SHA provides cryptographic confirmation only for artifacts with a declared SHA-256 pin in the manifest. “Pin is not declared” means “not verified against a known hash,” not that a file is automatically safe. Run custom .pt files only when you trust their source.

If a method does not start

On the model tab, press Check runtime; if it fails, open Dependencies... for ml.runtime.
Refresh file status, verify the actual directory and run SHA verification when a pin is available.
Select a downloaded model as active for the required type and retry from the CV workspace.
Inspect the reported error; Screph does not silently replace a missing model/backend with a cloud provider.

Runtime dependencies → · CV GUI and preview/apply →