Gpt-image-1 quality parameter

API
1

What really does the ‘Quality’ parameter mean here?

  1. Is it the sharpness of the image meaning high resolution?
  2. Or Is it the quality of rendering where in high quality, the output is better with more information with text having no typos?
2 Likes
2

In the gpt-image-1 Images API, “quality” does not change the image’s resolution (that’s what the size parameter is for). Instead, it controls the rendering fidelity—i.e. how much compute (and thus time and cost) the model devotes to each pixel, trading off speed versus detail and artifact suppression.

  • Quality = Fidelity, not Resolution.
    The dimensions of your output (e.g. 1024×1024, 1024×1536, etc.) are set by size. The quality setting—"low", "medium", or "high"—simply instructs the model how many sampling steps or how much internal processing to apply, which affects how crisp and detailed the final image looks).
  • Low / Medium / High:
    • Low: Fastest generation, lowest compute cost, coarser details, more artifacts.
    • Medium: Balance of speed and detail—good “everyday” choice.
    • High: Slowest, most compute-intensive, highest fidelity—sharper edges, finer textures, cleaner text rendering, and minimal artifacts.
  • Cost & Speed Implications:
    Higher quality settings consume more tokens (and thus cost more) and take longer to generate. In practice you might pay roughly $0.01, $0.05, or $0.20 per image for low, medium, and high quality respectively (for a standard square) depending on the exact pricing tier you’re on.

Bottom line: Use quality to dial in the level of detail and prompt adherence you need—particularly useful if you’re including small text or complex patterns—while size remains your control for actual pixel dimensions.

Hope this helps!

3 Likes
4

This makes sense, thank you!

2 Likes
5

No problem. We’ve got a few good gpt-image-1 threads here with more to come likely.

I’m still tinkering with locking down what I’m going to charge users for different quality settings, sizes, etc.

6

Great! One suggestion I have is the text rendering is completely faulty in ‘low’ setting. Either text rendering in low quality image can be completely disabled, or get it right even in low quality setting. I am not sure how expensive / complicated text rendering in image, given how text output llm’s seem to do it fairly cheap.

1 Like