Full resolution or nothing

Here's a dirty secret the AI image industry would prefer you never think about: almost every tool you've tried has been shrinking your photographs and betting you won't notice.

Not maliciously. Economically. AI inference cost scales brutally with resolution. Roughly quadratically, in fact, because pixels come in two dimensions: double the edge, quadruple the bill. So the standard industry move is simple. Downscale your image on the way in, process it small where the compute is cheap, then upscale the result on the way out with a sharpening pass to hide the crime scene.

The marketing page says "stunning high-resolution results." The pipeline says 1024 or 2048 pixels on the long edge. Your camera shot a detailed, textured master and the tool handed back a fraction of it, re-inflated like a supermarket tomato: bigger, redder, and tasting of nothing.

We think this is one of the great unexamined scandals of the AI tooling boom. And in real estate specifically, it's not a cosmetic problem. It's a commercial one.

Where premium campaigns actually live

Real estate media doesn't live on one screen, and it certainly doesn't live at portal resolution. The same master image runs as:

• a 200-pixel thumbnail in search results,
• a full-bleed spread in the property magazine,
• a full-size signboard on the front lawn,
• a floor-to-ceiling panel in a display suite,
• and a tight crop of the kitchen island in the social campaign.

The portals will crush your image down themselves. That's their job and they're welcome to it. Your master file's job is to survive everything the campaign throws at it for the life of the listing. A master is not a deliverable. It's an insurance policy.

Downscaled AI output voids that policy exactly where the premium money lives:

Print is unforgiving. A signboard produced from a 2K master goes soft from the footpath. The vendor who paid for the premium campaign walks past it every morning. They notice. Their neighbours notice. The next listing appointment on that street notices.

Crops are mandatory. No marketing team on earth uses the frame as shot. They crop for socials, crop for brochures, crop for the portal hero ratio. A crop of a downscaled image is a blur with a brand on it.

Texture is the luxury signal. Stone benchtops. Timber grain. Linen weave. Brushed brass tapware. The details that justify a price guide are precisely the frequencies a downscale/upscale round trip destroys first. Upscalers don't recover texture. They hallucinate a smooth approximation of it. Which means the "AI-enhanced" photo of the $40,000 marble island is showing a texture that marble doesn't have.

"But the upscaler is AI too!"

Yes. And think carefully about what that means. If a pipeline downscales your capture and then upscales it with a generative super-resolution model, then the fine detail in your delivered image was invented on the way back up. The grain in the oak floor is a statistical guess. The mortar lines in the brickwork are the model's best fiction.

We wrote a whole post about why invented pixels don't belong in listings. The downscale/upscale shuffle is the same sin wearing a lab coat: hallucination laundered through a resolution number. The metadata says 8000 pixels wide. The information in the file peaked at 2000.

Resolution is not pixel count. Resolution is information. You cannot upscale your way back to information you deleted.

What we do instead, and what it costs us

LUXengine processes your capture at up to a 4,000 pixel long edge and delivers at that resolution. No secret downscale to 2K. No generative upscale on the way back out. Drag RAWs into the platform and the finished JPEG that comes back is your photograph, edited, at working resolution. Run our Photoshop Workflow Plugin inside your own production pipeline and the output is TIFF, because production houses retouch, and nobody retouches a JPEG that's already been through someone else's compressor.

We would rather tell you the number than bury it, so here it is. 4,000 pixels is the current ceiling. Shoot bigger than that and we scale the file down to 4,000 and edit it there. You get back fewer pixels than you handed us, and we are not going to dress that up.

Here is why it is still the honest end of the trade: we scale down and then we stop. The pipelines this post is about scale down much further, edit at 1,024 or 2,048 where the compute is cheap, and then scale back up, so the file they return carries your original dimensions with invented detail poured into the gap. Same pixel count as your camera. A fraction of the information. And no way for you to tell which pixels are real.

Given the choice we would rather hand you 4,000 true pixels than 8,000 pixels that are mostly fiction. A 4,000 pixel master still holds a signboard, a magazine spread, a display-suite panel and a hard crop of the kitchen island. A 2K master re-inflated to look like an 8K master holds none of them, whatever the file properties claim.

And the gap in what the editing decisions are made against is not small. A pipeline working at 2,048 pixels sees roughly a quarter of the information we do; one working at 1,024 sees a sixteenth. Every call we make about a window pull or a blend seam is made against four to sixteen times more real detail, and nothing we hand back was hallucinated to fill a hole.

Raising the ceiling is an engineering problem we are working on, not a pricing tier we are holding back. When it moves, we will say so here.

We'll be honest about what this costs us too: it's harder engineering and more expensive compute, and we carry that instead of quietly passing the quality cost to you. It also puts a very healthy kind of pressure on our pipeline. Our editing decisions have to hold up at 100% zoom, because there's no upscaler downstream to smear over the mistakes. Every window pull, every blend seam, every colour transition gets delivered at the resolution where a professional retoucher will inspect it. That's the standard production houses hold humans to. We saw no reason to hold the machine to less.

The test we invite you to run today

Open any AI-edited image you've paid for this year. Put it at 100% and look at a textured surface: stone, timber, fabric, foliage. Then open the file that came out of your camera and look at the same surface.

One of two things is true. The detail survived, which means the tool actually processed something close to what you gave it. Or the texture went waxy, smeared and vaguely plastic, which means the file was shrunk, processed small, and re-inflated by a model guessing at what stone looks like. The pixel dimensions will not tell you which happened. Only the texture will, and that is the entire point of this post.

If it's the second one, you didn't buy an edit. You bought a preview of your own photograph.

Full resolution or nothing isn't a feature tier for us. There's no toggle and no "pro" plan where the detail comes back: every plan is processed the same way, at the same resolution, with nothing invented in either direction. We came from the end of this industry where the signboard goes up in the morning and the master file has to be good enough. Yours should be too.