Generating an image

From an empty field to a usable image, and why the third attempt is usually better than the first.

This is about Create in the Creative Studio, in the Image mode. Switch to Video and your face, product photos and scene come along; you do not have to pick them again. If you want fifty variants of one briefing, that is a campaign and not a single image; the steps below apply there just as well, but the button is somewhere else.

The short version

1. Decide what has to be in it before you pick a model.
2. Send your own images along as references, and mention them in the prompt.
3. Generate a test first (1K). Only when it is right, run it again as final.
4. If something goes wrong, change one thing and try again.

Steps 2 and 4 are where most of the credits can be saved.

With your credits, without your own account

If you have not linked your own OpenArt account, Denkploeg makes the image for you and you pay with your credits. You pick Fast, Standard or High quality and see the price up front on the card and on the button; you do not choose the model or the supplier behind it. Each image is charged separately. If it fails it costs nothing, and if only some of the images succeed you pay for those only. When your credits run out you see it straight away, with a button to buy more. If you do link your own OpenArt account, you work on that account as before. Images with a product photo or a brand model cannot use credits yet: make those with your own account.

Three quality tiers, and why that is the most expensive button

The same model costs well over twice as much at 4K as at 1K. Anyone who tries ten ideas at final pays hundreds of credits for nine images that get thrown away.

The tier controls the resolution and not the model. A test therefore shows you what you are going to get, just smaller.

Roughly what an image costs

Taken from the provider itself, in credits for one image in 4:5. This is not a price list but a ratio: it is about the order of magnitude.

Two things to remember. wan2-7-image charges a flat rate per image, so 4K is free there; that is the cheapest route to a large image. And gpt-image-2 on high costs six times a nano-banana-2 at 4K, so that is a choice for one hero shot and never for a series.

The screen tells you what it is going to cost before you press the button, and asks for confirmation. If it does not know the price, it refuses instead of guessing.

Which model for what

The list only shows models that can handle this job: the modes come from the provider's catalogue and the maximum number of reference images from the model's own input form. A model that cannot do it is not an option.

Watch the number of references. Models accept between three and sixteen. If you pick six product photos and a model that accepts three, you do not get an error but a silent cut-off: three photos disappear without anyone saying so. That is why the list filters on it.

References: the part that matters most

A reference image you send along gets a number. The first is @image1, the second @image2. The prompt has to refer to them, otherwise the model does with them whatever it sees fit.

This is the most common complaint ("it does nothing with my product photo") and the fix is one line. We write it in automatically:

Reference images: @image1 and @image2 show the model: keep the same face, hair
and build; @image3 shows the product: match it exactly in shape, colour,
material and detail

Every reference gets a role: model, product, background, anchor. That role is in the sentence above, and it is what tells the model that one is a face and the other is a jacket.

Prompting

The form the provider itself recommends, with the most important part first:

subject and action + place + composition + light + style

A short paragraph is enough. For a complicated scene it can be longer, in separate sentences per part.

Name a place, not a mood

This is the mistake that cost us a day, and it looks harmless.

Mood words give you a packshot: the product neatly in the middle, on an empty background. A concrete place gives you a scene. The model cannot draw a mood, only things.

Light: say where it comes from

Abstract terms like "soft lighting" and "diffused lighting" work poorly. Surroundings work: a time of day, the weather, indoors or outdoors, neon or studio light. If you say nothing about light, the model falls back on harsh flash, and everything looks as if it was shot with an on-camera flash.

Editing: say what stays the same

For an adjustment it works better to write down what must not change:

Change only the background to a plain gray studio wall. Keep the subject's face, hairstyle, clothing, pose, and lighting unchanged.

Without that list the model makes a new image that happens to resemble the old one.

And change one thing per attempt. Start with the biggest deviation: first the subject, then the composition, then the light, then the style. Changing two things at once means that when the result gets worse, you do not know which of the two did it.

The image never contains text

For us this is a rule and not a preference, for three reasons. Models render small letters unreliably. With a text layer, a forbidden term becomes an exact check instead of image recognition. And a second language then means redrawing instead of generating anew, with all the cost and all the uncertainty that brings.

If there does have to be text in the image itself, for example on packaging: put the exact words in quotation marks, add the position and the size, and use nano-banana-pro.

An image in another format

Cropping a 4:5 to 9:16 throws away exactly what carries the composition. grok-imagine-image-2 can reframe: it draws in the edges instead of cropping, and keeps the person, the clothing, the light and even the passers-by the same. That costs roughly thirty credits per shape, compared with a full regeneration that could just as well produce someone else.

Known problems

Having an image, clip or UGC video made in the conversation

The director in the conversation also makes images, video clips and (with UGC on) UGC videos, with the same engine as Make in Creative. Say what you want to see. It writes the brief, names the price in credits up front and only queues the image once you say you want it. Below its answer a card then shows the price and the button Sign and make (or Don't make). After your click it is made; the image or clip appears on the card and in the library. If you cannot sign, it waits under To sign for someone who can. A chat message never signs by itself. At most four images at a time. A UGC video is first saved as a draft and estimated; starting it needs the same signature, for exactly that amount. A clip is made on credits (Veo at Google), or with your own OpenArt account if you choose that.

What happens to the image

Every generated image goes into the library with its own identity: the provider, the model, the prompt fingerprint, the references that went along and what it cost. The failed ones too. That is not a hoarding habit but the answer to the question that will come in half a year: where does this image come from, and are we allowed to use it.

The provenance also goes into the file, as a C2PA manifest and in the IPTC fields. Platforms strip that on upload, so treat it as a courtesy to anyone who wants to check; the real record is in the database.

The media library

Under Media library in Creative Studio you find everything your workspace has in photos, videos and PDFs: what you uploaded yourself, the videos from Video Studio, what was made with AI and logos as raster images. It is the same picker as Make a post in Social: search by name, filter by type or source, a large preview, and in the Upload tab add photos, videos and PDFs by dragging, pasting or with the camera. You do not have to choose a type, but you first tick I am allowed to use these images: you confirm that you have the rights, and the file is added to the library as Owned by client. Without that tick nothing can be uploaded. A file from before this confirmation without established rights gets the chip Rights missing and you record them in the preview. SVG logos, fonts and the brand book belong to the brand profile and you upload them with Upload a brand file. In the ad editor you choose an image from the same library for each image slot with Choose another image. The same goes for Make (photos as reference or product photo, up to six) and for Design Studio under Image (an image to edit): you choose from the library or upload with the same picker, and an upload asks for the same confirmation.

See also

• The studios to see where this belongs
• Making a video
• Making and managing an avatar

Selected products appear by product name in the workbench. The reference sent to the image model remains internal.

New standalone generated images receive a name from the prompt and the Commercial licence classification. Existing files are not changed automatically. Library type groups can be collapsed; cards show image dimensions. Previews use a smaller file. Downloads offer format, width and compression while preserving the original. Trials continue to receive the limited watermarked version.