Layerize

layerize

Rebuild a source image as an editable CoDesign design: pixel layers from Seedream's layer decomposition, text rebuilt as real, editable text.

Inputs

• Source image — the file the user points at.
• Output path — where the .imgly deliverable goes. If the user named none, use the source's path with the extension replaced by .imgly.

Keep intermediate files (layer metadata, comparison renders, cut-outs) out of the output location.

Stage 1 — Layers (Seedream)

Call asset_generate — one successful call per image:

• model: bytedance/seedream-5-pro-layerize
• image_uris: [uri] — the source image, added to the workspace with asset_add({ source: { path } })
• An image under 512 × 512 pixels is refused: upscale it first (../models/upscalers.md), one more paid generation, and pass the upscaled result's id.
• prompt: exactly what the models skill prescribes for this model. Read ../models/bytedance-seedream.md before the call — it also explains the result (which asset is the background, why cut-outs come back rescaled, how bounding boxes are written).

Never retry a call that may have generated. If it times out, find its output with asset_search({ sourceId: 'ly.img.workspace.images' }) (newest first): the images created together with the source's dimensions or smaller, the full-size one being the background. If it returns an error instead, check the same way first; only when none of its output is in the workspace did it not run — fix the cause the error names and call it again.

After the third failure, stop: build nothing, report the errors with their gw_… request ids, and ask the user how to proceed.

Seedream is not optional: no other model and no local cut-outs replace it.

Save the returned metadata (layer names, z order, bounding boxes) as soon as you have it.

Place the layers with this edit code first — the page at the background's size, the background, then every cut-out at its bounding box. Paste the whole asset_generate result as RESULT:

jsCopy code
const result = RESULT;
const layers = result.metadata.find((m) => m.type === 'layers').data.layers;
const [bg, ...cutouts] = result.assets;
const width = Number(bg.meta.width);
const height = Number(bg.meta.height);
const [existing] = await engine.design.findByType('page');
const [scene] = await engine.design.findByType('scene');
const page =
  existing ??
  (await engine.design.create(
    { type: 'page', props: { width, height } },
    { parent: scene }
  ));
await engine.design.setProps(page, { width, height });
const image = (name, uri, [x0, y0, x1, y1]) =>
  engine.design.create(
    {
      type: 'graphic',
      name,
      props: {
        position: { x: x0, y: y0 },
        width: x1 - x0,
        height: y1 - y0,
        shape: { type: 'rect' },
        fill: { type: 'image', uri },
        contentFill: { mode: 'Contain' }
      }
    },
    { parent: page }
  );
await image('background', bg.id, [0, 0, width, height]);
for (const [i, asset] of cutouts.entries()) {
  const layer = layers[i + 1];
  const box = layer?.bounding_box?.absolute;
  if (box) await image(layer.name || `layer-${i + 1}`, asset.id, box);
}
return {
  type: 'text',
  text: JSON.stringify({ page, layers: layers.map((l) => l.name) })
};

Then handle Seedream's layers as follows (its output is not trustworthy as-is):

• Every layer sits at its bounding box; cut-outs come back rescaled, and contentFill: Contain fits each one into its box. If the bounding boxes are missing, find each layer's scale and position by matching it against the original.
• Drop Seedream's text layers; text is rebuilt in stage 2. Check that the background no longer shows the text. If it still does, inpaint it.
• Inventory: every visible element must be in exactly one layer.
• Drop duplicate layers.
• Cut any element Seedream missed straight from the original (take the alpha from the difference between the original and the composite).
• A part that sits in front of another object (e.g. a hand holding a product) may be its own layer, grouped with the object it belongs to.
• Background: keep the original pixels wherever no element covers them, and color-match the filled-in areas to the original.

Stage 2 — Text (you do the vision pass yourself)

For each visible text block, record:

• exact string with line breaks
• role (headline / sub / body / cta / legal / price)
• pixel bounding box and alignment
• cap height and line pitch in px
• font characteristics (serif/sans/script, width, weight, italic, caps)
• fill, stroke and shadow colors as hex
• whether it sits on a shape

Rules:

• Ignore text printed on products, packaging, labels, barcodes or logos.
• Keep a visually coherent block as ONE entry: a three-line headline is one block, not three.
• Transcribe literally; write "unknown" instead of guessing.

Set each block as real, editable text:

• closest Google Font (name a fallback too)
• uppercase through the text case property (text.case: 'Uppercase'), not typed capitals
• font size derived from the measured cap height
• letter spacing (text.letterSpacing, a factor) and line height (text.lineHeight: { visual }) fitted to the measured line widths and line pitch
• shadow (dropShadow) or stroke tuned to match the original
• a button or chip keeps its shape as a separate layer under the text

Matching the original beats CoDesign's typography lint: keep the original's tracking and leading even when the lint warns about them.

Build

• One page at the source's pixel size.
• Layers from back to front: background, scene groups, logo, text.
• Give every layer a descriptive name (background, left-figure, product-can, logo-…, headline-…).
• Group each scene's layers together, and group the text blocks.

Acceptance

Compare against the original: place the source image once more as the top layer, at the page's size, with blend: { mode: 'Difference' }, and preview it. Where the rebuild matches, the preview is black; every bright shape is a difference — a missing or doubled element, text off its baseline, a wrong color. Then destroy that layer.

Check with a PNG export (export({ format: 'png' })) compared against the original. Fix and re-export until every point holds:

• text baselines within ±2 px, line widths within ±2 %
• text shadow visually matches the original
• average absolute pixel difference below 10/255
• no element missing, no element appearing twice
• named layers, grouped per scene

Then run the judge loop against the source (../judge/SKILL.md), export with export({ format: 'imgly', outPath: '<output path>' }), and report:

• the layer list
• what Seedream got wrong and how it was fixed
• where the intermediate files are