How to Edit Product Photos With Krea 3: Number Your References, Name What Stays, Count the Cost

Krea's new image editor works from up to eight reference pictures. A routine for getting the edit you meant, and the pricing quirk that should shape it.

Krea 3's price list has a quirk worth noticing before you open it. At 2K, the finished picture costs six cents. Each 2K reference picture you send along with it costs seven.

That one line tells you how the tool wants to be used. Krea 3, which opened in beta on October 15, is an editor built around references: you give it a photo, point at other pictures by their position in the list, and describe the change. It can take up to eight. The obvious way to use it is to throw everything you have at it. The good way is almost the opposite.

My position: the routine that gets you the edit you meant is the same routine that keeps the bill small. Fewer references, sent smaller, described clearly, with the parts that must not change named outright. Here is that routine, drawn from Krea's own documentation and Comfy's guide to the tool.

What Krea 3 does

Krea 3 is an instruction-based image editor. You start from a picture, add up to eight supporting images (a product, a person, a location, a material), and write what you want: "put the sneaker from the second image on the person in the first image." That example comes straight from Krea's API documentation. It returns a 1K or 2K image. With no input pictures at all, it works as a plain text-to-image model.

The feature that matters for product and commercial work is regional editing. Krea's product page puts it simply: "Draw a box and say what changes. Krea 3 redraws inside it and leaves the rest of the image as it was." In ComfyUI the same switch is called precise_edit. For a packshot where the label, the logo and the lighting on the bottle must stay exactly as shot, that is the whole game.

You can use it in the Krea web app, through Krea's API (currently early access for beta workspaces), on Replicate, or as a ComfyUI Partner Node billed in Comfy credits.

The cost math that should shape your routine

From Krea's API documentation:

• A 1K output costs $0.03. A 2K output costs $0.06.
• Each 1K reference adds $0.02. Each 2K reference adds $0.07.
• The slower "high" thinking mode, which plans complicated edits more carefully, adds $0.02.

So the same job can cost very different amounts:

• A 2K edit from your photo plus one 1K reference, in the default fast mode: $0.08.
• A 2K edit with four 2K references: $0.34.
• A 2K composite with eight 2K references in high mode: $0.64.

None of these will bankrupt anyone on a single image. Across a catalogue of two hundred products, each edited three or four times, the gap between the first and last example is real money. And as the next section shows, the expensive version is often not even the better one.

The routine

1. Send references at about one megapixel

Krea's docs are direct about this: "Do not upscale a reference before you send it." Around one megapixel (roughly 1000 by 1000 pixels) is enough for most edits. Anything over four megapixels gets shrunk to four anyway. Upscaling a reference adds cost and time and gives the model nothing new to work from.

2. Use four images or fewer

This is the step that saves the most. With five or more references, Krea shrinks each one to at most 1536 pixels on its longest side. Fine detail suffers first, and on product work, fine detail is usually the label text. Krea's own advice is to use four or fewer when detail matters. If you have eight candidate references, pick the four that each do a different job.

3. Put the picture you are editing first

In ComfyUI the references connect as image_1 through image_8, numbered in the order you connect them, and when the aspect ratio is set to automatic, the output follows the first image. Putting your base photo first keeps the frame you shot. The same habit is worth keeping in the Krea app, so your numbering always means the same thing.

4. Point by position and by contents

Comfy's guide says to refer to each image by its order. Krea's docs say to refer to each image "by what it shows." Do both: "the green bottle in the second image," not "image 2" alone, and not "the bottle" alone. In ComfyUI's node you can write @Image2 and still add what it shows. Saying both gives the model two ways to find what you mean.

5. Name what must stay, not only what changes

Comfy's Krea 3 guide makes this point, and it is the one most people skip. "Change the background to a marble counter" leaves the model free to tidy up the label too. "Change the background to a marble counter. Keep the bottle, its label text, its cap and its reflections exactly as they are" does not. Write the keep list like a retoucher's note.

6. Do local edits before global ones

Krea separates two kinds of edit. Regional edits redraw only what needs to change. Regular edits, meaning restyles, relighting and reframing, redraw the whole picture. The docs warn: "Regular edits re-render the whole image, so artifacts can build up when you keep editing the latest output." So do the careful, local work first, on a clean image. Save the whole-image changes for last, and do as few of them as you can.

7. When a chain drifts, branch from the best step

Comfy's guide suggests restarting from the strongest earlier result if a sequence of edits drifts. Keep every intermediate image and name it with a step number. When step five looks worse than step three, go back to three and try again from there, rather than pushing on and stacking more damage.

8. Check the change, not just the picture

Put the edit on a layer over the original in your editor, set the blend mode to Difference, and anything that turned non-black has changed. On a regional edit, the change should sit inside the area you asked for. This takes thirty seconds and catches the label that shifted two pixels before the client does.

A worked example

Say you shot a green glass bottle on a white sweep, and the client wants it on a marble kitchen counter by a window, with a sprig of rosemary beside it.

Your references, in order: the packshot first, a phone photo of the marble counter second, a still of window light you like third, a rosemary sprig fourth. Four images, each about one megapixel. Fast mode, 2K output, regional editing on.

The instruction, written like a retoucher's note:

"Place the green bottle from the first image on the marble counter from the second image. Light it with the soft window light from the third image, coming from the left. Add the rosemary sprig from the fourth image lying beside the bottle. Keep the bottle, its label text, its cap and its highlights exactly as they are in the first image."

At Krea's list prices that is $0.06 for the 2K output plus four 1K references at $0.02 each: $0.14 a try, assuming each one-megapixel reference bills at the 1K rate. One caution: ComfyUI's docs say its node scales each reference to the output size before upload, so at 2K there the references may bill at the 2K rate, which would make it $0.34. Check your first receipt. Run it three times, keep the best, and check the label with the Difference blend before anything else.

Where it breaks

You cannot train it on your own products yet. Some launch coverage, including a sponsored article on TestingCatalog, described Krea 3 as an editor you can fine-tune with your own add-ons. Krea's own API documentation says otherwise: sending add-on files returns an error that reads "LoRAs are coming soon to the API." Style references are not available yet either. Plan for the tool as it exists, not as it was announced.

Same seed, different picture. The seed is the number that normally lets you repeat a result. Krea rewrites your prompt behind the scenes before generating, so, as the ComfyUI docs note, the same seed will not reproduce the same image. Save the outputs you like; you cannot regenerate them on demand.

The speed and price comparisons are Krea's. The product page says a 2K edit finishes in under eight seconds, but a footnote calls that an illustrative playback time, and the comparison clips are "pre-generated outputs, not a live benchmark." Its "3× cheaper" headline sits beside prices of about $0.06 for Krea 3 and $0.135 for Nano Banana Pro, which works out nearer 2.25 times. Treat the numbers as claims to test.

ComfyUI pricing is not published yet. The Partner Node runs on Comfy credits, but Comfy's pricing page lists only the older Krea 2 models at the time of writing. Check your balance after a test run before you queue a batch.

Exact pixel sizes are not guaranteed. You can request a width and height, but Krea's docs say the output snaps to the nearest supported ratio and to a 1K or 2K canvas. If a placement needs exact dimensions, plan to crop afterward.

Read the licensing page before client work. Krea's product page footer links to a "Model Licensing" page, and its contact-sales list mentions intellectual property protection for enterprise customers. Neither says, on the pages I checked, what applies to an ordinary paid account. Check before you deliver.

What to take from this

Most AI editing advice tells you to give the model more. More references, more detail, a higher setting. Krea 3's pricing and its own documentation point the other way. Four references beat eight. One megapixel beats an upscale. A keep list beats a hopeful prompt. A local edit beats a global one.

Try it on one product this week. Run the same edit twice: once with every reference you have at full size, once with the routine above. Compare the label text, then compare the bill. If the cheap one is also the better one, you have your house style.