Where does the machine help, and where should a person decide?
The AI look reads as a mistake when nobody chose it. These are design choices, offered as Metaesthetic's working rules rather than research findings:
- Machine for the background, camera for the subject. A generated setting behind a photographed person makes the contrast look intended.
- Real typography. Set the type yourself. Generated lettering is one of the functional tells on the signature page.
- Leave one flaw in. An asymmetric detail or an imperfect edge shows that a person looked at the frame.
- Make the glow a system. If every image in a set shares the same treatment, it reads as a house style.
- Keep the provenance. If you publish generated images, keep their Content Credentials intact; these platforms read them to label posts.
What canvas can each model give you?
From each vendor's own documentation, checked 13 November 2026.
| Setting | OpenAI GPT Image (gpt-image-2.5 Sunburst and Flare) | Google Nano Banana 2.1 (gemini-nano-banana-2.1) |
|---|---|---|
| Sizes offered | Recommended: 1024 x 1024, 1536 x 1024, 1024 x 1536; plus custom sizes | 1K, 2K or 4K for each aspect ratio |
| Aspect ratios | Any, from 1:3 to 3:1 | 14 fixed: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 1:4, 4:1, 1:8, 8:1 |
| Custom size rules | Width and height in multiples of 16; no edge over 3,840 pixels; 655,360 to 8,294,400 pixels in total | Not offered; pick a ratio and a size tier |
| Largest output | 8,294,400 pixels (3840 x 2160); sizes above 2560 x 1440 are experimental | 4K tier, for example 4096 x 4096 at 1:1 or 5504 x 3072 at 16:9 |
| Default | Size, quality and background accept "auto" | Matches the input image's size, otherwise a 1:1 square; Gemini 3 image models default to 1K |
| Quality settings | low, medium, high, xhigh, max, auto | Not a parameter; set the size tier instead |
| Provenance | Checked by OpenAI's provenance tool for C2PA Content Credentials and SynthID | SynthID watermark in every generated image |
The full Nano Banana size table, and the pixels needed for common print sizes, are on the print size planner.
Sources
Checked 13 November 2026:
- OpenAI: Image generation guide, Content provenance guide
- Google AI for Developers: Nano Banana image generation, Gemini API
Next step
When the default look is the problem, see breaking it for the controls that move output away from it.