Image review

Optionalplugin: image

The image review plugin shows a round of images an agent generated, such as illustrations, product shots or icons, and asks for a verdict on each. You look at every image closely, choose the one to take forward, and mark the parts to change on the image itself. The agent gets each mark in coordinates an image-editing or inpainting step can use.

Click to enlarge

The favourite, with a box and a pin on the parts to change and a note on the image as a whole.

When to use it

• Illustrations and spot art, where you choose one direction from several.
• Product shots and marketing images, where details need fixing before they ship.
• Icon sets and app artwork, checked on light and dark backgrounds.

Install

The plugin comes with the app. Install it in Settings › Plugins, or from the command line:

pinrail plugins install image

What you see

• The round’s notes from the agent, across the top, in view while you work.
• A list of the images, each with its size, format and verdict. A review of a single image shows the image alone.
• Each image large on a stage, to zoom with the scroll wheel or = and -, and to pan by dragging with the space bar held. Fit and 1:1 return to the whole image and to its own pixels.
• A backdrop behind transparent pixels: a checkerboard, light or dark. The agent picks the first one, and G switches.
• The agent’s reasoning, the prompt it used, and how the image was made, such as the model and the seed.
• Favourite, Keep or Drop for each image, and a note on it if you add one. There is at most one favourite. A review of a single image offers only Keep and Drop.
• Regions to change. Drag a box over part of an image, or click a spot, and say what should change there.
• The previous round. When the agent sends a new round, C switches to the same image from the round before, with the regions you marked on it.

Press ? for the keys.

To see it before any agent asks with it, send its sample: pinrail submit image --sample, or Send a sample in its details in Settings › Plugins.

Asking from your agent

AGENTS.md

## Before using generated images

When you generate images for me to choose from, submit them to Pinrail with
the `image` plugin and wait for my decision:

1. Write the payload: a `notes` line on what this round is, and one entry per
   image with a stable `id`, a `name`, its `file`, and your `reasoning`.
2. Run: `pinrail submit image --title "<what the images are for>" --data images.json --attach <each file> --wait`
3. Take the `favorite` forward. Keep the `keep` images as alternatives, and
   drop the rest.
4. Apply each region's `note` to the area it marks, and each image's `note`
   to the image as a whole, then submit the next round with `--revises <id>`,
   keeping each image's `id`.
5. If the command exits 5, stop and use none of the images.

What the agent sends

images.json

{
  "notes": "Round 1 of the empty inbox illustration: four directions.",
  "subject": { "name": "Tern empty-inbox illustration", "use": "320 px wide, light and dark" },
  "backdrop": "checker",
  "images": [
    { "id": "A", "name": "Paper plane",
      "file": { "$attachment": "paper-plane.png" },
      "reasoning": "Mail as something that **flies off and is done**.",
      "prompt": "Flat vector spot illustration, a paper plane looping over hills",
      "details": { "model": "flux-1.1-pro", "seed": 48211 } }
  ]
}

• Each image is a file sent with --attach and named in file: PNG, JPEG, WebP or SVG, up to 25 MB each and 16 in a round.
• notes and reasoning are markdown. details are short facts, shown as chips.
• backdrop is the first backdrop behind transparent pixels: checker, light or dark.
• favorite says whether the person can mark one image as their favourite. The default is true when the review has more than one image, and false when it has only one. Set it to false when each image only needs to be kept or dropped.

What comes back

{
  "decisions": [
    { "id": "A", "action": "favorite", "note": "Keep the palette exactly as it is",
      "size": { "width": 960, "height": 720 },
      "regions": [
        { "shape": "box", "x": 0.9083, "y": 0.4111, "width": 0.0917, "height": 0.1389,
          "px": { "x": 872, "y": 296, "width": 88, "height": 100 },
          "note": "Remove this cloud: it is cut off by the edge" }
      ] },
    { "id": "B", "action": "keep" },
    { "id": "C", "action": "drop", "note": "Too dark for the light theme" }
  ],
  "undecided": ["D"]
}

Field	What the agent does with it
decisions	Takes the favorite forward, keeps the keep images as alternatives, and sets the drop images aside. A note applies to the image as a whole.
regions	Changes each marked area as its note says. x, y, width and height are fractions of the image from its top left, and px is the same area in the image’s pixels. A box can serve as an inpainting mask. A point marks a spot.
undecided	Images you gave no verdict. An image with regions and no verdict counts as kept.

Reference

The plugin’s manifest, and the schemas a payload and a decision are checked against, read from the plugin’s own files.

Manifestmanifest.jsonPayloadwhat the agent sendsDecisionwhat comes back

FieldsJSON

name	image
version	1.0.0
title	Image review
description	A round of generated images or illustrations, sent as PNG, JPEG, WebP or SVG files, each large on a stage to zoom and pan. Pick a favourite, keep or drop the rest, and box or pin a spot on an image to ask for a change there.
use_when	You generated candidate images or illustrations (a spot illustration, a hero, a product shot, an icon set) and need a person to pick one, shortlist or drop the rest, and mark the areas to fix, remove or redraw, in coordinates an image-editing or inpainting step can act on.
attachments	.png .jpg .jpeg .webp .svg image/png image/jpeg image/webp image/svg+xml · up to 25 MB each, 16 a review
shortcuts	JNext image KPrevious image PPin a comment at the centre of the view BBox what is in view (zoom in first) =Zoom in -Zoom out 0Fit the image 1Actual pixels arrowleftPan; the other arrows likewise HDrag to pan rather than to mark GNext backdrop: checker, light, dark CCompare with the previous round FMake this image the favourite SKeep this image on the shortlist XDrop this image

• notesstring

  Markdown: what this round is, and what to look at.

  at most 10000 chars

• subjectobjectrequired

  • namestringrequired

    What the images are of, e.g. Tern empty-inbox illustration.

    1–120 chars

  • usestring

    Where the image will be seen, e.g. 320 px wide in the empty inbox, light and dark.

    at most 200 chars

• favoriteboolean

  Whether the person can mark one image as their favourite. The default is true when the review has more than one image, and false when it has only one. Set it to false when each image only needs to be kept or dropped.

• backdropenum

  What shows through transparent pixels at first. Default checker; the reviewer can switch.

  "checker" "light" "dark"

• imagesobject[]required

  1–16 items

  • idstringrequired

    1–40 chars

  • namestringrequired

    1–80 chars

  • filefilerequired

    The image as a file sent with --attach: PNG, JPEG, WebP or SVG. An SVG is shown as a picture; its scripts and links never run.

  • reasoningstring

    Markdown: the idea behind this one, and what to look for.

    at most 4000 chars

  • promptstring

    The prompt it was made from, shown as it is.

    at most 8000 chars

  • detailsobject

    Short facts about how it was made, shown as chips: model, seed, steps, style.

  • altstring

    What the image shows, for a screen reader.

    at most 300 chars

• decisionsobject[]required

  A verdict per image the reviewer decided. At most one favourite.

  • idstringrequired
  • actionenumrequired

    favorite: the one to take forward. keep: still on the shortlist. drop: set it aside.

    "favorite" "keep" "drop"

  • notestring

    What to change in the image as a whole, or why it was dropped.

    at most 4000 chars

  • sizeobject

    The image's size in pixels (an SVG's in viewBox units): what the regions' px are measured in.

    • widthnumberrequired
    • heightnumberrequired
  • regionsregion[]

    Changes asked for on parts of the image, each on a box or a point.

    • shapeenumrequired

      box: an area to change. point: a spot to change.

      "box" "point"

    • xnumberrequired

      The box's left edge, or the point, as a fraction of the image's width from its left.

      0–1

    • ynumberrequired

      The box's top edge, or the point, as a fraction of the image's height from its top.

      0–1

    • widthnumber

      The box's width, as a fraction of the image's width.

      0–1

    • heightnumber

      The box's height, as a fraction of the image's height.

      0–1

    • pxobjectrequired

      The same region in the image's own pixels, origin top left. For an SVG, in the units of its viewBox.

      • xnumberrequired
      • ynumberrequired
      • widthnumber

        at least 0

      • heightnumber

        at least 0

    • notestringrequired

      What to change there.

      1–2000 chars

• undecidedstring[]required

  Images left without a verdict.