What does this AI photo generator actually do?
You type a prompt, pick a model, pick a size, and get 1 to 4 images back. There is no canvas, no layers and no brush. The form is the whole interface.
Three things are worth knowing before your first run:
- You choose the model, per image. The models are not tiers of one product. They fail in different ways, and the cheap one is the right answer more often than people expect.
- 9:16 is offered first. Everything else on this site renders vertical, so the image sizes start where the videos end up. Other aspect ratios are in the same chip row.
- Reference images are a model feature, not a site feature. Nano Banana 2 takes up to 14 of them, GPT Image 2.5 takes one, and Z-Image takes none.
Open the photo form and the model chips sit above the prompt box. If you already know which one you want, go straight in: Z-Image, Nano Banana 2 or GPT Image 2.5.
The images come out as stills. This surface never returns a video. If motion is what you need, that is the AI influencer generator or the video tools.
Which model should you use?
Start from the job, not from the model name. The decision table first, the specifications second.
| The job | Model | Credits at 1K | Why this one |
|---|
| Background plate for a slideshow or an ad | Z-Image | 1 | Nobody looks at the plate. Pay the least. |
| Bulk concepts, 20 ideas in one sitting | Z-Image | 1 | Cheap enough to throw most away |
| Your product, kept recognizable | Nano Banana 2 | 2 | Up to 14 reference images hold the shape and the color |
| A face you need again next week | Nano Banana 2 | 2 | References keep the same person across runs |
| A packshot with a readable label | GPT Image 2.5 | 2 | It is the strongest of the three at text inside the frame |
| A thumbnail with two words on it | GPT Image 2.5 | 2 | Same reason: legible words, correctly spelled |
| A busy scene with many objects | Nano Banana 2 | 2 | Long prompts, up to 4,000 characters, get followed |
| Model | Credits 1K / 2K / 4K | Reference images | Prompt ceiling | Typical wait |
|---|
| Z-Image | 1, 1K only | None, text prompt only | 900 characters | 5 to 15 seconds |
| Nano Banana 2 | 2 / 3 / 4 | Up to 14 | 4,000 characters | 10 to 30 seconds |
| GPT Image 2.5 | 2 / 2 / 4 | 1 | No stated limit | 20 to 60 seconds |
Nano Banana 2 is Google's image model and GPT Image 2.5 is OpenAI's image model. Z-Image runs here as a model choice with no vendor claim attached.
Two things fall straight out of that table. GPT Image 2.5 costs the same at 2K as at 1K, so if you picked it, take the 2K. And Z-Image has no 2K or 4K at all, so anything you will crop into or print needs one of the other two.
How do you write a prompt that returns a usable product shot?
A weak product prompt names the product. A strong one names the photograph.
Write in this order, and the hit rate climbs immediately:
- Subject. What the thing is, in plain words. "A matte black 500ml water bottle."
- Surface and setting. "On a wet concrete ledge, morning, city street behind."
- Light. The single most useful clause in the prompt. "Hard side light from the left, deep shadow on the right."
- Lens language. "Shot on a 50mm lens, shallow depth of field, shot from slightly below."
- Framing for the crop. "Centered, space above the bottle for a headline."
- What to exclude. "No hands, no text, no logo."
Two habits that fix most bad results:
Describe the light, not the mood. "Cinematic" and "premium" are not instructions. "One softbox at 45 degrees, dark background, no fill" is.
Leave room for the overlay. If the image is going under a slideshow headline or above a caption, say where the empty space goes. Recropping a full frame costs another credit.
Prompt ceilings matter here. Z-Image stops at 900 characters, which is about six of those clauses, so keep it tight or move the shot to Nano Banana 2 with its 4,000-character ceiling. For the packshot version of this workflow, the AI product photo generator opens on the same form with product defaults.
What does a reference image actually change?
A prompt describes a category. A reference image pins an instance. That is the whole difference, and it decides which model you pick.
What references are good for:
- Your actual product. The real bottle, the real label, the real color. A prompt cannot recover a shape it has never seen.
- One person across many images. Upload the same face and the creator stays the same creator. That is what the AI influencer generator uses to keep a persona consistent.
- A style you already own. Feed back an image you liked and ask for a different angle of it.
- A set. Six references of the same object from different sides gives the model the geometry, not a guess at it.
What references do not fix: a bad prompt. The reference tells the model what the object is, and the prompt still has to say what the photograph is.
The practical limits are the model limits. Nano Banana 2 takes up to 14 reference images, which is enough for an object and a person and a set in one run: start at the Nano Banana 2 image generator. GPT Image 2.5 takes exactly one, so choose it when the reference is the label rather than the object. Z-Image takes none, which is the honest reason it costs 1 credit.
Why is the cheap model the right default?
Most images on a content site are not the subject. They are what the subject sits on.
A slideshow background, a gradient behind a hook, a texture under a chat screen, a scene you will crop to 20% of frame: none of those reward 4 credits. They reward volume, because the tenth attempt is usually better than the first, and at 1 credit you can afford a tenth attempt.
Spend the extra credit when one of these is true:
- A human face is the subject, and it has to survive being looked at.
- A real product has to stay itself, which needs references.
- Words have to be legible inside the image, which is GPT Image 2.5.
- You need 2K or 4K, which Z-Image does not offer.
A working split for a week of content: plates and concepts on Z-Image, hero shots and people on Nano Banana 2, anything with type on GPT Image 2.5.
Run the arithmetic once and the habit sticks. Forty background plates on Z-Image is 40 credits. The same forty on the 4K tier of Nano Banana 2 is 160 credits, which is almost a quarter of the $49.99 plan, spent on images nobody will look at directly.
What are these images for on this site?
This is not a general art tool bolted on. Each of the four uses feeds another surface.
| Use | Best model | Where it goes next |
|---|
| Ad creative and product shots | Nano Banana 2 for the product, GPT Image 2.5 for the label | Cut into a conversation ad or a demo clip |
| Slideshow backgrounds | Z-Image at 1 credit | Behind the hook slide in the slideshow generator |
| AI creator portraits | Nano Banana 2 at 2 credits, Z-Image at 1 | The persona portrait in the AI influencer generator |
| Thumbnails and covers | GPT Image 2.5 | YouTube, TikTok covers, anywhere words sit in the frame |
The creator portrait is the one that pays for itself. Two credits buys a face you then reuse across a whole campaign, and the same face in every ad is what makes a set of ads look like one brand instead of five stock photos.
For user-generated-content style shots, where the product looks like somebody's kitchen counter rather than a studio, use the UGC photo generator. The prompt rules invert there: ask for phone camera light, slight clutter and an off-center crop.
What does an image cost?
Priced per image, by model and resolution. Nothing is priced per prompt or per attempt, so a run of four images costs four times one image.
| Model | 1K | 2K | 4K |
|---|
| Z-Image | 1 credit | Not offered | Not offered |
| Nano Banana 2 | 2 credits | 3 credits | 4 credits |
| GPT Image 2.5 | 2 credits | 2 credits | 4 credits |
In dollars, on the $49.99 plan for 700 credits, a credit is about $0.0714:
- A Z-Image plate: about 7 cents.
- A Nano Banana 2 hero at 1K: about 14 cents. At 4K, about 29 cents.
- A run of four GPT Image 2.5 images at 2K: 8 credits, about 57 cents.
Plans, read from pricing on September 30, 2026: $9.99 for a one-time 100-credit pack, $49.99 per month for 700 credits, $199.99 per month for 3,000 credits. On the larger plan a credit is about $0.067, so the same four-image run is about 54 cents.
New accounts start with zero credits, so the first image needs a plan. Cancel anytime, 14-day refund window on unused credits.
What the photo generator does not do
Short list, and it matters more than the feature list.
- No editor. No layers, no masking, no inpainting, no brush and no object removal. You change the prompt and run it again.
- No video. This surface returns stills. Motion lives on the AI influencer generator and the video tools, and nothing here exports 16:9 or long form.
- No scheduling, no auto-posting, no connected accounts. You download the file and post it yourself.
- No ad account integration. Images do not flow into Meta or TikTok Ads Manager. You upload them there.
- No API and no batch import. One run at a time in the browser, 1 to 4 images per run.
- No analytics. Nothing here tells you which image performed. That data stays in your ad platform.
- No guarantee of a rights-cleared output. Do not prompt for a real person, a brand mark you do not own, or a character somebody else owns.
If an editor is what you actually need, get the generation here and do the retouching in the tool you already have. The alternatives comparisons list what the neighboring tools cover.
