Ming Design-Layer splits a design into RGBA layers: Sume does not

Sume has no layer-decomposition endpoint, so it cannot do what Ming-Image-0.1-Design-Layer does: split one flattened design into several RGBA layers. What it can do is generate each element as its own transparent PNG with background: "transparent" on GPT Image 2.5.

Ming facts are from its Hugging Face card; Sume facts from the Image API docs. Both read 2026-10-01.

What does Ming Design-Layer do?

Per its model card, it takes a flattened design image plus a layer plan (a prompt, or a requested layer count) and returns RGBA PNG layers. It is a 6B model under the MIT license, recommends 1024 (or 512) input, 12 steps and CFG 2.0, and expects an 80 GB GPU.

Does Sume return layers?

Not as a documented feature. The Image API docs describe text-to-image and reference-guided editing that return finished images; layers are not part of the request or response. Two earlier posts cover the same gap for Qwen layered output and Seedream layer decomposition.

How do I get separate elements?

Generate them one at a time. background accepts auto, transparent or opaque on GPT Image 2.5, and transparent needs a format that carries alpha, so ask for png. You then composite the PNGs in your own editor or renderer.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "A single red kettle icon, centered, no shadow, no text",
    "image_size": "1024x1024",
    "background": "transparent",
    "output_format": "png"
  }'

What if I already have a flat image?

Sume's background removal tool takes a public HTTPS image_url and returns one subject cutout. That is one cut, not a stack of layers, and it does not recover what was hidden behind overlapping objects.