Anand Geetha

Card 23 of 38· Domain 3 · Computer vision

Generating and editing images and video

Which model and which call, why there is never a URL to read back, and the mask rule that is inverted from most people’s intuition.

Generating and editing images and video
Open the card in a new tab to read it at full size.

Generation is a small topic with a lot of specific constants, and two of them behave in ways that catch people out.

Which model, which call

Task Model or filter The call
Text to image Filter the catalogue by text to image. GPT-image-1, -mini, -1.5, -2, FLUX, MAI-Image client.images.generate
Editing or inpainting The same GPT-image family images.edit, as multipart/form-data
Text to video Sora 2 — filter by video generation client.videos.create, asynchronous

Reading the image result

GPT-image always returns base64, in result.data[0].b64_json.

response_format is not supported. There is no URL option at all. You decode the bytes and write the file yourself.

This matters because plenty of image APIs elsewhere hand you a URL, so a URL-returning answer looks entirely plausible. It is wrong here every time.

Image parameters

  • size
  • quality — low, medium or high
  • n
  • output_format — PNG or JPEG. WEBP is not supported.
  • background=transparent requires PNG.

Video mechanics

Video is asynchronous: videos.create, then poll videos.retrieve, then videos.download_content.

  • seconds accepts 4, 8 or 12.
  • remix applies one targeted edit.
  • input_reference drives image-to-video.

The mask rule

Masks are inverted relative to most people's intuition, so read this twice.

  • The mask is a PNG.
  • Fully transparent pixels — alpha 0 — mark the region to edit.
  • The mask must match the image dimensions exactly.

So the transparent part is what gets repainted, not the opaque part. Most people assume the reverse: that they are painting over the area they want changed. Getting it backwards produces an image where everything except the intended region has been regenerated, which is a memorable way to learn it.

Safety and provenance

  • Every generated image carries C2PA Content Credentials — provenance metadata recording that it was AI-generated.
  • Photorealistic images of minors are blocked by default.
  • A harmful prompt returns no image, rather than a filtered or degraded one.

Retired, and never the answer

dall-e-3 was retired on 4 March 2026, and dall-e-2 before it. Neither is ever a correct answer now.

Their old parameters travel with them as distractors: quality=hd/standard and style=vivid/natural belong to DALL-E, not to the GPT-image family. Seeing either in an option is a reliable signal that the option was written for the previous generation.