GPT Image 2.5 transparent PNG: fix opaque backgrounds

For a transparent GPT Image 2.5 image, request background: "transparent" and output_format: "png" or "webp", then check the original file’s alpha values. A PNG extension, a gray checkerboard and the presence of an alpha channel each tell you less than whether the background pixels are actually transparent.

This guide is for developers and designers preparing product cutouts or reusable image assets. The settings follow the OpenAI image generation guide, checked September 9, 2026, for both Flare and Sunburst. The file checks below run locally; they are not a benchmark of either model’s cutout quality.

First distinguish a viewer grid from a painted grid

Do not use a screenshot as the test file. A screenshot usually captures the viewer’s rendered background along with the image. Save the generated image itself, and keep a copy before sending it through an editor or optimizer.

Request a format that can preserve alpha

For generation, the relevant arguments are:

result = client.images.generate(
    model="gpt-image-2.5-flare",
    prompt="A ceramic cup isolated on a transparent background, no text",
    size="1024x1024",
    background="transparent",
    output_format="png",
)

This fragment assumes an initialized OpenAI client. The complete API guide includes setup, credentials and saving b64_json to a file. Calling the model is billable; checking an existing file is not.

For an existing product photo, use images.edit() and pass the reference image. Ask to remove the background while preserving the product’s shape, colors and label. Inspect those details afterward: preserving identity is an instruction, not a guarantee that the result is identical outside the background.

For transparency, both PNG and WebP are suitable formats. JPEG has no alpha channel. Renaming product.jpg to product.png does not convert the file or recover lost transparency. Also check whether the application wrapping the API actually exposes and forwards the background option.

Inspect alpha with a local script

Install Pillow in your Python environment with python -m pip install Pillow. Save the script as check_alpha.py:

import sys
from PIL import Image

with Image.open(sys.argv[1]) as source:
    print("format:", source.format, "mode:", source.mode)
    print("dimensions:", source.size)
    alpha = source.convert("RGBA").getchannel("A")
    histogram = alpha.histogram()
    total = sum(histogram)
    print("fully transparent pixels:", histogram[0])
    print("partially transparent pixels:", sum(histogram[1:255]))
    print("opaque pixels:", histogram[255])
    print("non-opaque fraction:", round(1 - histogram[255] / total, 6))

Run it against the original download:

python check_alpha.py product.png

Converting to RGBA for inspection also handles palette-based transparency. If every pixel is opaque, conversion creates an opaque alpha channel; it does not remove the background. If only a few pixels are non-opaque, the image still may not be a useful cutout. Inspect where those pixels are located.

Check the edges, not just the background

Place the original over a dark fill and a light fill in an editor. Look for pale halos, missing handles, damaged label lettering and shadows that end abruptly. Inspect at the actual delivery size as well as zoomed in: resampling can change the appearance of a clean edge.

If the background is opaque, correct the format and background arguments first. If the background is transparent but the product has changed, refine the editing instruction or return to the original reference. Repeatedly editing the latest output can make it harder to identify where a detail was lost.

For files that must remain pixel-identical inside the product, use a workflow that explicitly preserves those source pixels and validate that requirement separately. Generative editing should not be described as lossless background removal.

Keep transparency through delivery

Export to an alpha-capable format with background flattening disabled. Open the exported file again and rerun the alpha check. If a downstream tool only accepts JPEG, agree on an intentional solid background; there is no transparent JPEG option to enable.

The size guide helps separate generation dimensions from final export dimensions. Use the model comparison when deciding which model to try, and the pricing guide before running repeated edits.