Image

Draw labelled rectangles onto an image. The visualization half of vision-grounding loops: render a document or frame, let a model propose bounding boxes, draw them, let it look again.

Why use this

• Visualization, not mutation — shapes are drawn as given and clipped at the image edges; a box that hangs off the canvas renders partially, because showing a wrong proposal is the point of a review loop.
• One coordinate space with the pdf module — pixels, top-left origin, matching what Pdf.Rasterizer reports and Pdf.FormFields consumes, so boxes flow between rendering, preview, and field placement untranslated.
• Bytes in, bytes out — buffered Uint8Array payloads (PNG, JPEG, or WebP in; the same set out, chosen via format), composing with S3.Get/S3.Put, Octet.Decoder, and HTTP bodies without touching the filesystem. Each output reports its mediaType.

Kinds

Kind	Purpose
Image.Blank	Produce a solid-color canvas as image bytes (png/jpeg/webp) — pipeline seed or hermetic test fixture.
Image.Overlay	Draw labelled rectangles onto an image; returns annotated bytes (png/jpeg/webp) plus dimensions.

Example

kind: Telo.Application
metadata: { name: box-preview, version: 1.0.0 }
imports:
  Image: std/image@latest
---
kind: Image.Overlay
metadata: { name: DrawBoxes }
stroke: { color: "#FF3B30", width: 3 }
label: { color: "#FFFFFF", placement: top-left }
---
# Draw the model's proposed fields onto the rendered page.
- name: marked
  inputs:
    image: "${{ steps.page.result.image }}"
    shapes: |-
      ${{ inputs.fields.map(f, {
        "x": f.x, "y": f.y, "width": f.width, "height": f.height,
        "label": f.name + " (" + f.type + ")"
      }) }}
  invoke: !ref DrawBoxes