Overview
GPT Image 2 is OpenAI's image model (gpt-image-2), and its standout trait is that it actually listens. Built instruction-first on an autoregressive approach, it follows your description closely and nails layout, composition, and the little details.
The real surprise is how accurately it renders text inside images — including Chinese, Japanese, and Korean. Poster headlines and cover taglines come out crisp and readable instead of garbled.
Capabilities
It handles both text-to-image and image-to-image. You can feed it reference or input images (up to about 16, in jpeg, png, or webp) so it creates or edits based on your own material.
Resolutions come in three tiers: 1K, 2K, and 4K. Editing is a real strength here — paired with a reference image, it makes precise, controllable changes rather than wandering off on its own.
How to use here
Here, just write the scene you want, pick a resolution, and generate.
For touch-ups, upload your reference or the image you want changed, describe the tweak, and it will follow your lead.
Credits
You pay per image, and higher resolution costs more: 1K is 5 credits, 2K is 10, and 4K is 20.
As a rough guide, 1 credit is about ¥0.1, so you can keep a loose sense of the cost.
Best for & tips
Reach for it when you need accurate in-image text, strict prompt adherence, or precise reference-guided edits — think posters, captioned covers, UI mockups, and brand assets.
If you just want cheap images in bulk, Seedream 5.0 Lite is the more economical pick. For multilingual text or intricate layouts, start at 2K for steadier results.