Transparent Backgrounds with GPT Image 2

Cutout alpha is a first-class parameter on GPT Image 2. Set output_format to png, set background to transparent, and the model returns a clean cutout without you having to run a background removal pass.

The minimal call

1const cutout = await fal.subscribe("fal-ai/gpt-image-2", {
2  input: {
3    prompt: "A single matte black wireless speaker, three quarter angle, cleanly cut out.",
4    image_size: "1024x1024",
5    quality: "high",
6    background: "transparent",
7    output_format: "png",
8  },
9});

The returned PNG has a proper alpha channel. Drop it on any background without halos or fringing.

When to use it

• Product shots you plan to composite onto marketing backdrops.
• App icons where you need a transparent canvas.
• Logos and marks where the surround is supposed to be the customer's page.
• UI stickers and emoji that need to drop into an interface.

What it costs extra

Nothing. Transparent output is billed at the same tier as opaque. The model runs the same sampling budget and returns the alpha as part of the image.

When it will not give you clean alpha

If the prompt describes a product with soft edges (fur, smoke, a glass surface) the alpha will include partial transparency at the edge, which is correct but can look weird over a colored background. Test your composite against the actual deployment background before you ship.