What 3 AI Images for a Website Really Cost
Real prices from a real account: what one Nano Banana 2 image costs at 1K, 2K and 4K, the exact prompts, and the mistake that put two of me in the same picture.
Almost nobody publishes real numbers for AI images. You get “affordable” and “starting from,” and then a pricing page in credits that means nothing until you have already paid.
So here are mine. I needed three pictures for the front page of this site. This is what they cost, the prompts I used, and the mistake that cost me four extra credits.
Every number here comes from my own account on Aug 7 2026.
The prices
I checked these against the API before generating anything, so they are the real charge, not an estimate.
Model: Nano Banana 2 (Google), on Higgsfield.
| Size | Credits per image |
|---|---|
| 1K | 1.5 |
| 2K | 2 |
| 4K | 3 |
To turn credits into money you need the plan price. I am on Higgsfield Plus: $49 a month for 1,000 credits. That makes one credit about 4.9 cents.
| Size | Cost per image |
|---|---|
| 1K | ~7 cents |
| 2K | ~10 cents |
| 4K | ~15 cents |
I used 2K for all of them. 4K is not worth 50% more for a web page, because the picture gets shrunk anyway. 1K is too soft once it fills half a screen on a laptop.
What the job actually cost
| Credits | |
|---|---|
| First 3 images | 6 |
| Re-doing the one that failed (2 attempts) | 4 |
| Total | 10 |
Ten credits. About 98 cents. For three finished pictures on a live website.
The re-shoot was 40% of the bill. That is the part worth paying attention to, because it was avoidable.
The mistake: two of me in one picture
I asked for a simple scene — one person at a desk, filming himself with a phone on a tripod. I wrote this in the prompt:
Camera: wide three-quarter view from slightly behind and to the side, showing both him and the phone he is talking to.
Here is what came back.
Two of him. Sitting at the same desk.
The room is perfect. The light is right. The character matches. And it is unusable, because there are two people in a picture about one person working alone.
Why it happened. I asked for a shot from behind and to the side that showed both him and the phone. To the model, “behind” plus “and both” reads as two subjects. So it drew two. It was doing what I asked; I asked badly.
The fix
Two changes.
1. Say the number of people out loud, at the top.
CRITICAL: there must be EXACTLY ONE PERSON in the entire image. One single man, alone in the room. No second person, no duplicate of him, no twin, no mirror, no reflection, no one else anywhere in frame.
2. Pick a camera angle that cannot imply a second person. I replaced the over-the-shoulder framing with a plain side view:
Camera: a clean side-on view from the left, roughly at desk height, so we see the man in profile on the right and the phone on its tripod on the left, with clear empty space between them.
That worked first time.
Note the small detail the model added on its own: his face on the phone screen. I never asked for it. That is the difference between a model guessing and a model understanding the scene.
Three things that saved money
Check the price before you generate. Most APIs will quote a job without running it. On Higgsfield it is one flag, get_cost. Free, instant, and it stops you finding out afterwards.
Attach a reference image every single time. All three pictures use the same character because I attached the same reference file to each prompt. Without it, the face, hair and clothes drift between images and the set falls apart. Generating four pictures that do not match is four wasted charges.
Ban text in the prompt. I put a line like this in every one:
Absolutely no text, no words, no letters, no numbers, no logos anywhere.
Image models still render text badly. One of my scenes was a wall calendar, and without that line it would have filled itself with broken fake dates. Broken text is the fastest way to make a page look cheap.
One more number: file size
The three images came out of the generator as PNG files totalling 16.9 MB. Putting those straight on a website would be slow, especially for someone on a phone.
Converted to WebP at quality 82–84:
| Size | |
|---|---|
| Original PNG (3 files) | 16.9 MB |
| WebP (3 files) | 243 KB |
Ninety-eight percent smaller. I could not see a difference on screen. This step costs nothing and takes a few seconds, and skipping it would have hurt the page far more than the pictures helped it.
The short version
- Nano Banana 2: 1.5 credits at 1K, 2 at 2K, 3 at 4K.
- On a $49 / 1,000-credit plan, a 2K image is about 10 cents.
- Three finished website images cost me 10 credits, about 98 cents — including one mistake.
- Asking for a shot from behind that shows both the subject and what he faces will get you two subjects. Say “exactly one person” and pick a side-on angle.
- Attach a character reference to every prompt, ban text in every prompt, and convert to WebP before it touches your site.
The pictures from this job are live on the front page of this site.