Build a product catalog that stays consistent

One clean mask does not make a clean catalog. Crop, scale, position, shadow, and spacing must match too.

Split the job in two: segmentation finds the subject; composition places it.

Start with a measurable target

Define the output before you run the batch:

• Canvas dimensions and aspect ratio.
• Background color or transparency.
• Minimum margin around the subject.
• Preferred visual scale.
• Horizontal and vertical alignment.
• Shadow style, if any.
• Output format, quality, and naming convention.

Try a 1600×1600 white canvas with the subject centered at 85–90% of the safe area. Consistency matters more than the exact number.

Keep the mask lossless

Do not treat a compressed preview as the source mask. Resizing and lossy encoding damage edges. Keep one lossless mask at source resolution.

Fix that mask once, then export every format from it.

Review by category, not only by average quality

Average scores hide failures. Review by category:

Category	Typical failure
Hair and fur	Chopped strands, solid-looking edges
White products	Missing low-contrast boundaries
Glass	Removed reflections or opaque interiors
Thin structures	Lost straps, spokes, cables, stems
Dark-on-dark	Background leaking into the subject
Soft shadows	Shadow removed when it should be composed

If a model raises the average but breaks jewelry chains, do not ship it.

Correct exceptions without breaking the batch

Create one project per image. Store correction strokes in normalized coordinates and apply them to the original mask at export.

You get three things:

1. Corrections stay aligned at different output sizes.
2. The original image never changes.
3. Every export is reproducible from the project revision and recipe.

Treat composition as versioned data

Instead of “use whatever settings Maria used last Tuesday,” keep a recipe:

{
  "version": 1,
  "background": { "type": "color", "color": "#ffffff" },
  "crop": true,
  "margin": 64,
  "width": 1600,
  "height": 1600,
  "scale": 0.9,
  "positionX": 0.5,
  "positionY": 0.5,
  "format": "jpeg",
  "quality": 92
}

Months later, the same recipe can render new and old products alike. A new look gets a new version.

Use fair scheduling for shared systems

Do not let 1,000 files create 1,000 active GPU tasks. Process one image per workspace, then rotate. Big batches move without blocking everyone else.

Preserve the handoff

Keep the handoff flexible:

• Transparent PNG for design and compositing.
• Transparent WebP for the storefront.
• White JPEG for marketplace ingestion.
• Color and alpha files for automated pipelines.
• Layered PSD for manual art direction.

Build an asset that survives the whole workflow.

A small release checklist

• Inspect representative categories at 100% zoom.
• Confirm dimensions, margins, position, and background color.
• Validate transparency and file signatures.
• Check that metadata is stripped where expected.
• Test partial batch failure and retry behavior.
• Confirm outputs expire on schedule after downstream ingestion.
• Record model, pipeline, recipe, and dataset versions.

Try blankground, explore batches, or connect the API.