Link to the cookbook:
Here are a few samples from my first tests:
Finally! There are many of us that has been waiting for this​
Touché… well played… it’s about time… 
Just so you know: I think the record player is on fire - thank me later 
(the boy being chased is a really nice vivid demo, thanks!)
This didn’t need a meandering cookbook. One liner:
"background":"transparent" is no longer refused by the API with gpt-image-2.
Use my existing API AI image editor, where I did not gate the background values by model, to make a “generate” request.
Done:
Save to PNG, WebP, or GIF file format supporting the received transparency
I note an odd thin grey border, likely being a “color key” halo for the API backend method used to programmatically prompt and make transparent imagery. The RGB behind transparency actually contains a grey border larger than the alpha channel, and then more imagery beyond a border.
I would need to change the workflow to show more of the underlying “RGB leak” to you, as the image support library will flatten out pure macroblocks.
Yay, anywhere that you follow the use cases of the first post here, to create image assets to overlay somewhere, will contaminate the whole of the other imagery with “It’s AI!” watermark detections.
Actually, that was an image edit → transparency.
Not quite sure what you are talking about. The image was verified.
What am I talking about?
OpenAI uses a SynthID variant to embed AI image identification watermarking into the surface of output imagery.
Cookbook’s suggestion:
Design-template assets: Generate app icons, stickers, and decorative elements that users can place on any template or background.
You have a beautiful hand-painted background. You make an in-game sprite as suggested in the video, or other API image, transparent or not, put them in your “seasonal background”. End users can “validate” that everything you are delivering as a final image product says “Generated with OpenAI tools”.
Not your company, not your product, not the user’s product, not the core of your creative content as final product - but something for the general public to now hate on wholesale with the watermark inclusion (intrusion) creating “flagging” saying “OpenAI”. Will your picture of a pillow with AI art be called “FAKE! AI SCAM STORE”?
Sorry, but I quess I’m just simple-minded…
Nice, now we can finally have proper pets!
BUG
The PNG I requested and got back from my first trial at quality:low has a flaw in the API.
The imagery is not truly opaque where it should be.
The alpha channel values of the car are 253 of 255, with some subtle noisy variations being alpha = 252-254. I checked again with “medium” and 1536x1024 generations - the opaque areas still are not 255 = solid, they are 252-254, the majority 253.
This means that in composing and compositing, other extremes of black and white will still show through. The alpha channel in API images delivered will need to be normalized or clipped.
An unrequested border halo symptom remains - showing here on my white background image viewer destination, when using a transparent background and prompting a white border from API:
Keying out the white makes this more dramatic:
To counter the current symptoms, new client software feature will need to be another “transparent halo removal”, one that acts on and expands existing alpha transparency (instead of being informed by key color in expanding for its own transparent conversions.)
Finally, this is pretty useful. I’ve been waiting for proper transparency support in the API, especially for assets that need to be reused across different backgrounds.
The cookbook examples look handy too. I’m kinda curious how clean the edges are with things like hair, glass, or other semi transparent parts. Those are usually where background removal gets messy.
Has anyone tested those kinds of images yet?
(post deleted by author)
You can have antialiasing in imagery at the same time as you have a continuous alpha channel. (I now answer a deleted post).
Besides the smooth edges with no color other than that of imagery, Tux here will cast an alpha channel shadow on a light background.
The AI’s generation of a “transparent color” is absolute and is a “cut out”, however.
In this picture of dice, see the checkerboard background through the true transparency, then sent for “edit” (where transparent 100% is no longer a mask redraw area).
Prompt: These dice I’m sending as input have full RGBA 32 bit transparency transmitted to your API. Change the color of the foreground die from red to orange – preserving the same see-through appearance for whatever background the resulting image is composed upon.
The AI “transparent background” image no longer has the checkerboard canvas showing through.
Additionally, besides the dominant 253 value alpha, enjoy the underlying RGB image contents returned by the API from gpt-image-2 once alpha is discarded:
A technique of “blendables” that falls down when prompting for white against the transparency.