Lithi Image Analysis

Start with one bounded image observation

For a synthetic 32 × 32 image, ask for the foreground square and return its top-left-origin bounds. Keep image identity, source dimensions and coordinate convention with the result. An empty observation, unreadable image and ambiguous observation are different outcomes that remain visible to the reviewer.

Your agent defines what to inspect in a permitted image and how to handle ambiguity. It prepares the source identities, dimensions and coordinate rules. Lithi returns bounded observations or explicit exceptions. Your agent checks them against the original image before making any decision that depends on them.

Your agent keeps the observation question and final decision. The Kit prepares the supported image request; participating Macs process the bounded images, and your agent checks coordinates, identity and ambiguity against the source.

Images, scans, screenshots, and image regions belong to the Documents & Images information family. Decide what the model may observe or classify before sending images. The operation defines the observation and evidence it returns.

This guide is the public workflow for image analysis. Vision uses the universal Compute skill; it has no separate lithi-vision skill. Keep policy decisions and external actions outside the image task.

Use only images you are permitted to process under the selected data profile. Do not treat the filename, output order, or a model label as a durable identity.

Conceptual task brief: Inspect images IMG-01–IMG-03 for the declared foreground region. Preserve each image ID, source dimensions, and top-left coordinate system. Keep unreadable, ambiguous, and empty observations distinct. Return bounded regions or explicit exceptions with their source identity. Accept only when every image is reconciled and each region maps back to the original source.

The selected image operation and supported integration map that brief to the permitted request and returned evidence. It does not identify a person or authorize a consequential action; the customer remains responsible for the decision that follows the observation.

Preserve source and region identity

Assign every image a stable item ID. When a result names a region, preserve the coordinate system, source dimensions, and region identity needed to map that observation back to the original image. Tiling or resizing changes the evidence boundary; record how regions map back to the original source.

Decide whether the result must include image-level labels, bounded regions or boxes, confidence or review state where the operation defines it, citations to image or region identity, and exception states for unsupported or unreadable files. A vision output is evidence for the declared observation. It is not permission to identify a person, make a consequential decision, or trigger an external action unless a separate authorized system explicitly owns that decision.

Fit the visual budget and review exceptions

Check image count, encoded bytes, dimensions, and any visual or token budget in the current operation limits. Do not silently lower fidelity when the requested decision requires details the transformed image no longer contains.

Unreadable, unsupported, ambiguous, and review-required images remain explicit. Reconcile them by stable item identity. A synthetic worked image can teach the review process, but it does not prove live model quality or capacity.

Worked example

Inspect a synthetic image batch

Create a 32 × 32 grayscale PNG with a black square on white. The declared observation is the foreground square. Its expected top-left-origin bounds are x = 8 through 23 and y = 8 through 23, or normalized x = 0.25, y = 0.25, width = 0.5, height = 0.5.

These are fixture geometry, not measured detector output. An all-white image may produce an empty observation; that is different from an unreadable file. Preserve image identity, coordinate convention, and source dimensions in either case. A detected region remains evidence for the declared observation only; it does not authorize identifying a person or taking an external action.

Compare published prices

You can see what it costs to analyze images with an AI model and with an image service, using each provider's published list price. These tables compare published prices for the same work, not output quality.

The example labels and describes 5,000 keyframe images:

• Each image is 1,024 × 1,024 pixels.
• Each image gets one pass.

Each AI model counts image tokens with its own rule, and image services bill per image.

Published list prices checked 5 October 2026.

Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows with a note count input with that provider's own published rule. Each service bills in its own unit. Lithi figures use the published example metering profile; a signed quote sets the real charge. These tables compare published prices for the same work, not output quality.

Use Documents & Search when image regions sit inside document evidence and Audio & Video when frames or media timing are part of the request. Use Workflow verification to choose checks and First test for a bounded sample.