Markup and project

Selector workspace

• Canvas shows the current image and annotations.
• Toolbar toggles selection, creation, moving, resizing and linking.
• The properties panel shows the type, name, geometry, descriptions and relationships of the selected object.
• Tree and graph help to find objects and check their relationships.

Creating and editing elements

1. Select the creation tool and draw the object’s region on the canvas.
2. Provide a type and friendly name, then add a description if necessary.
3. Use the moving and resizing tools to refine the boundary of the object.
4. Test the object on the desired frames or source states.

Selecting multiple objects allows you to perform supported common operations. Before making a bulk change, save the project or make sure the action can be undone.

Features and regions of interest

A feature describes a visual property of an element or its fragment. It can be used as a template, area for a computer vision method, or material for comparison.

• The feature is associated with the owner and the original image.
• Crop, mask and preview help you check the saved material.
• Copying a feature creates separate linked files if required by the selected operation.
• The region of interest limits the measurement or processing to a portion of the image.

Relationships and groups

Element relations connect elements and features, for example by specifying a hierarchy, owner, or domain relation. Selecting a relation shows its type and linked objects.

Groups are saved in the project and created by supported workflows. There is no general visual editor for custom groups in the current version.

Text and voice descriptions

The description can be associated with an element, feature, relationship, match, or series of images. The text is edited before saving to the project.

Voice annotation uses the selected speech recognition method. With local Vosk the audio remains on the computer; with the cloud option, the recording is transferred to the selected service.

Multiple views and image analysis

Canvas layouts allow you to see different materials of the same project at the same time. Zoom, scrolling, and cursor can be synchronized between views.

• Magnifying glass and pixel view show values under the cursor.
• The histogram and line profile help evaluate brightness and color.
• Layers and comparison allow you to compare the original image with the result of the method.

The presentation layout is a local setting and may not transfer with the project.

History of changes

Supported markup changes can be undone and redone. Once you open another project or end a session, the available history may change, so use separate saved copies for important steps.

Saving and Loading

• Ctrl+S: save the current project.
• Ctrl+O: open Project Manager and select a project or autosave.
• Save As creates a project in a new location. Afterwards, check the linked images and results.

When saving, the project structure and the availability of associated files are checked. A warning about a missing image or result must be corrected before transferring the project to another user.

Project manager and automatic saving

The manager lists discovered projects, autosaves and files that cannot be read. Before restoring an autosave, make a copy of the current project.

After duplicating, moving, or renaming a project, open it from the new location and save again to check all the content references.

Computer vision

Computer vision operations create results separately from the main annotation. Running, previewing, sequences, graphs, and applying results are described in the computer vision manual.

Screph Code

Screph Code is the built-in agentic IDE for working with files and requests based on project data. After saving the project, you can transfer selected images, annotation, and authorized material to a separate code workspace folder. The Screph project and that folder are saved separately.

• Builder can respond with text and change files in the selected working folder.
• Before prompting, save your working folder to version control or as a backup.
• After the answer, check the modified, created and deleted files.
• Builder can modify files and execute commands within the selected access mode. Stop ends current work after resource cleanup, but does not undo files already written. An agent response does not establish that the final program works.

Review source files, attachments, and edits in the files, objects, and changes tabs. When leaving an unsaved file, you can save it, discard changes, or cancel navigation. If a file changes externally, review it before replacing the text in the editor.

The model and Code engine are selected separately. By connection, Codex, and OpenHands are available when their connection and environment are prepared. Sharing the chat appearance with Agent does not merge history, context, or permissions. Codex Desktop is a separate application; passing a draft to it does not submit a request to the model.

Screph Code chat

Datasets and model training

1. Create a dataset and set its name, task, and classes. Use Add data to import files, a folder, selected Capture frames, or Canvas material. General includes Flowers mini, a sample dataset that you can edit like any other dataset.
2. Assign classes and create bounding boxes or masks on the canvas. For an image with no objects, confirm an empty scene; missing annotations do not count as a negative example. Exclude incorrect samples and review the quality of unreviewed ones.
3. Resolve the data issues listed under Next step. Choose the model to train, weights, device, and parameters. If the environment is not ready, Start training opens its setup window; the dataset has already been saved by this point.
4. The first Start training checks inputs, freezes a dataset version, and creates a preprocessing preview using real samples. Confirm the run separately after reviewing the preview. Dataset editing is locked during preparation and training; Stop cancels current work.
5. Results shows history, metrics, run comparisons, and errors. Make corrections in an editable dataset and check them in a new run; old versions and results are not overwritten. Repeating an identical successful run is blocked until data or settings change.
6. In the simple workflow, promote the candidate with validation-split evaluation. Advanced view also offers a separate final evaluation on the held-out test split. Prediction preview does not replace that evaluation.
7. On the Model page, release a named model, choose the primary model, and complete its usage profile. To pass it to the final program, enable the separate permission and specify the input source and how the result will be used. Releasing a model or choosing it as primary does not automatically enable transfer to the agent.

The Datasets and training tab uses the same canvas. Classification assigns one class to an image, detection stores object boxes, and instance segmentation stores separate binary masks. Simple and Advanced show the same data, settings, and results; the program mode restricts available tasks and models.

A Canvas region is linked to a dataset sample through explicit import. Moving or deleting the region does not by itself change dataset annotation already saved. Images from one sequence and crops from one source keep a common group when splitting data, so adjacent frames do not enter both training and independent validation.

Selecting a thumbnail once does not replace the canvas image. Open selected on Canvas opens a sample for annotation; double-click opens a separate preview. For detection, choose a class and ML · Rectangle, then draw around the object; you can adjust or delete the box or assign another class.

For segmentation, use the polygon, brush, and eraser tools. Create a separate mask for each instance. Point and box prompts require MobileSAM to be prepared; propagation to neighboring frames uses the selected available processor. Accept, reject, or correct each proposed mask; it does not automatically become confirmed annotation.

Needs review, Confirm annotation, No objects in the image, and Exclude from training have different meanings. An excluded sample is retained but does not participate in training. After restoring a sample, review its quality again, even if its previous class is retained.

In the training preview, inspect transformations of real images and annotations: size, cropping, padding, and augmentations. Cancel returns to settings without starting training. Switching tabs retains the pending preview; confirm it after returning to Learning.

Add current image/annotation explicitly imports the selected material. In classification, a region becomes a separate crop; in detection and segmentation, the full frame is retained with its box or mask. For quick import, use Create → Region → dataset and choose the dataset and class above the canvas.

Agent opens a read-only Learning assistant with training and validation summaries and the current configuration. Assign its model separately on the LLM Models for tasks page. Images, paths, and the held-out test split are not sent to this chat; it cannot start training or edit data. Annotation and augmentation by the dataset assistant are separate jobs with review and explicit acceptance; accepted proposals still require quality review.

Saving training and recovery

By default, long local training runs independently of the application window. Reopening reconnects history to the saved run. If process termination is confirmed and a complete epoch checkpoint is available, explicit resume is offered. An unknown state does not authorize automatically starting another copy of training.

The draft, frozen dataset version, run, candidate, and promoted model are saved as distinct states. Use the provided project, dataset, or model export commands for transfer; a weights file alone does not carry annotation and history. COCO export is available for detection and segmentation, and Screph-manifest YOLO for detection.

Camera calibration and frame use

In Capture → Source, Camera calibration… is available for camera, stream, and file sources in General, Industrial, and UAV. Select the lens model and target, and confirm the same camera, size, focus, and zoom, with no prior lens correction.

Frames can be added from files, the timeline, or automatic collection from an already running source. Calculation requires at least eight fitting frames and three held-out validation frames. After Calculate, review errors and warnings; save the profile as JSON and the session as ZIP. The ZIP is saved separately from the project and lets you continue later.

Correct image… creates a result for the same camera and image size. Use new validation frames after comparing models. Calibration, a tutorial model, and successful image processing do not establish camera, robot, or flight control; spatial scale is verified separately.

Before shutting down

1. Save the project and wait until the files are written.
2. Check warnings for related content.
3. Create a backup copy of the project directory and, separately, the working folder with the code.