It is easy to pick the highest image quality setting and leave it there. With GPT Image 2 on PixelDojo, that habit can use ten times as many credits as starting at Low. The published prices for standard-size images are 0.5 credits for Low, 1.5 for Medium and 5 for High. These are PixelDojo credits, separate from OpenAI’s direct API pricing.
The sourdough infographic comparison offers a useful example: one prompt, one title, five labels and four numbered steps, generated at each quality level. The question is how much of the extra detail your audience will notice, and whether the cheaper version already does the job.
Myth 1: Low quality means broken text
Low does not necessarily mean unreadable. In the sourdough example, the Low version reproduced ANATOMY OF A SOURDOUGH LOAF, all five labels and all four steps correctly. That included “Bulk ferment 5 hours” and “Bake at 250°C”. The layout was simpler and the textures flatter, but the words were there.
One successful image cannot tell you how reliably the setting will perform across a campaign. It does give you a reason to try it. If you need thirty social graphics or ten poster headlines, make a small test using your actual copy before committing to the whole batch. A simple design may need nothing more.
Test difficult cases separately. A bilingual shop sign and an app screen crowded with small labels present different challenges. Check every character, number and punctuation mark before increasing the volume. An impressive High example tells you little about how often Low will reproduce your particular wording correctly.
Myth 2: Medium is just a slightly sharper Low
The Medium sourdough example added explanatory paragraphs beneath the labels and steps. The page looked fuller, but some of that fullness came from copy that had not been requested.
That does not mean Medium always adds text. It is an observation about this particular output. A new generation can alter the layout or introduce details at any setting.
For a mood board, extra wording might not matter. For a client infographic with approved copy, it means another review and perhaps another render. Include a clear instruction: “no additional text beyond the quoted strings”.
Then read the result. The instruction gives the model a firmer brief, but it does not remove the need to proofread. Check for invented claims as well as spelling mistakes.
Myth 3: High is always the right call for final work
High is worth comparing when texture, fine lines or realistic lighting matter. PixelDojo lists High as the maximum-detail setting for GPT Image 2, costing 5 credits at standard sizes. In the sourdough comparison, that version added a framed border and finely hatched step icons.
Whether those details justify the cost depends on where the image appears. Fine hatching can disappear in a small social post. In a full-page illustration, readers have more room to see it.
For detailed photography or artwork, compare High with the cheaper result at the size people will actually view. For a graphic built around a short headline, start by checking Low and Medium. If the words are clear and the design works, paying for more detail may add little.
Inspect the final file before deciding. The setting is a useful starting point, but it cannot tell you whether a particular image meets your brief.
Myth 4: 4K is just an upscale
GPT Image 2 supports generation at large output dimensions, and PixelDojo lists a 3840×2160 option. Its published 4K price is 10 credits at High or 3 at Medium. Choose the output size as well as the quality setting; they affect different parts of the job.
A 4K image is not automatically suitable for every print size. At 300 pixels per inch, 3840×2160 gives you roughly 32.5×18.3 centimetres before cropping. That is smaller than A2. The dimensions you need depend on the physical print size, viewing distance and your printer’s requirements.
Check the downloaded file’s actual dimensions. If you regenerate at another size, review the design again too. Line breaks, spacing and the positions of small labels can change. Approval of a smaller version does not replace checking the final export.
A cheaper workflow that holds up
For a poster, begin with four Low variations. At the standard prices above, those cost 2 credits in total. Choose the strongest direction and make one Medium version for another 1.5 credits. Look closely at the copy, spacing and anything the model has added.
If you then need a standard-size High render, the total comes to 8.5 credits. Two standard-size High renders would cost 10 credits.
A High 4K final changes the arithmetic. Four standard-size Low variations, one standard-size Medium version and one High 4K final total 13.5 credits. That is less than two High 4K renders at 20 credits, but more than two standard-size High images.
Reusing the same prompt will not necessarily preserve the layout you liked. Use the chosen image as a reference when making the next edit, and check what changes.
Here is a prompt to try:
“Swiss style concert flyer printed on cream stock. Headline: ‘LOW TIDE SESSIONS’ in heavy condensed sans serif. Subhead: ‘Three nights of ambient music on the pier’. Date block: ‘4 TO 6 JULY’. Callouts: ‘Free entry’ and ‘Bring a blanket’. Every word spelled exactly as written. No additional text beyond the quoted strings.”
The honest caveats
Check the current charge before running a batch, because prices and available model options can change. Compare alternatives using the same brief and intended output size. A lower price per generation only helps if enough of those images are usable.
Choose the file format for its destination. PNG preserves lettering without lossy compression. JPEG or WebP can make web images smaller, although you should check that compression has not damaged small text. For print, ask about dimensions, colour handling and delivery format instead of assuming that a PNG is ready to send.
Finally, identify each piece of copy as a headline, subhead, callout or button. Those roles help communicate the hierarchy you want. Keep wording short where you can, review it at the final display size, and increase the quality setting when there is a visible problem for it to solve.