Common AI Image Artefacts and What Causes Them

Every AI image tool produces artefacts sometimes. Every one. The tools that claim otherwise are the ones whose outputs you should check most carefully.

An artefact is anything in a generated image that a real camera pointed at a real scene would not have produced — a soft halo around an edge, a badge that has gone slightly molten, a straight line that bends. Most AI car photo marketing pretends these do not exist. We think the opposite approach serves dealers better: name every failure mode, explain what causes it, and tell you which ones actually matter for a listing.

This is the field guide.

1. Boundary halos and edge fringing

What it looks like: a faint glow, light rim, or soft fringe along the edge of the car where it meets the background — most visible along the roofline against the wall, or around door mirrors.

What causes it: the boundary between car and scene is where generation does its hardest work. The model is making per-pixel decisions about where your car ends and its generated scene begins, and blending the two so the join is invisible. When that blend is slightly too generous, a few pixels of transition become visible as a halo. Fine structures — aerials, mirror stalks, roof rails — give the model the least information to work with, which is why fringing clusters around them.

Does it matter? Usually invisible at thumbnail size. On the full-size image, a noticeable halo reads as "edited photo" and is worth a re-run.

2. Melted detail: badges, grilles, wheel spokes

What it looks like: small, intricate structures losing their crispness. A badge whose lettering has gone soft. A mesh grille where the pattern smears. Wheel spokes that blur into each other or — the classic — a wheel that gains or loses a spoke.

What causes it: generative models reproduce fine repetitive detail less reliably than broad shapes. They work from learned patterns, and at the scale of a badge or a spoke pattern, "approximately right" and "exactly right" diverge. The model knows what alloy wheels look like in general; it has to reproduce your wheel exactly, and small dense detail is where that is hardest.

Does it matter? Yes — this is the category to take most seriously, because it touches the car itself rather than the scene. A buyer will not notice one spoke at thumbnail size, but badges and wheels are exactly what enthusiast buyers zoom in on. Check them against your original on every output. If a badge or wheel has degraded, do not publish that image — re-run it or use the original photo. The legal dimension of accuracy in listing photos is covered in what AI car photo editing actually changes.

3. Geometry warps: bending panel lines and straight edges

What it looks like: a shut line that drifts, a window edge with a subtle curve it should not have, a sill line that wobbles where it crosses from car into shadow.

What causes it: generative models have no engineering drawing of your car — they produce images that are statistically plausible, and "plausible" tolerates small geometric drift that manufacturing does not. The risk concentrates where a line on the car passes through a region the model is reworking heavily: the boundary zone, dark areas, or surfaces where reflections are being regenerated.

Does it matter? Mostly invisible at listing size, but the human eye is unusually good at straightness, so a warp big enough to see is big enough to register as wrong. Scan the long lines — roofline, shoulder line, sills, window frames — when you check the output.

4. Texture repetition in generated scenes

What it looks like: the background giving itself away — floor tiles that repeat a little too perfectly, a wall texture with a visible rhythm, identical scuffs appearing twice.

What causes it: the scene around the car is fully generated, and when a model produces large areas of uniform texture, its learned patterns can recur at visible intervals. Real floors have irregular wear; generated floors sometimes have suspiciously regular character.

Does it matter? Rarely. It sits in the part of the image buyers look at least, and well-designed studio scenes — simple, soft, low-detail — give the model little opportunity to repeat itself. This is one reason clean studio environments produce more reliable output than busy photorealistic locations.

5. Reflection and lighting mismatch

What it looks like: paint reflecting a sky that is not in the scene, glass showing the old forecourt, a car lit warm in a cool-lit studio, a shadow disagreeing with the light.

What causes it: your original photo arrives with its old environment baked into every reflective surface, and the model must reconcile that with the new scene it is generating. This is why Motuva renders reflections to match the new scene rather than leaving the old ones in place. It is the deepest technical challenge in the whole category — deep enough that we wrote two dedicated pieces on it: why AI car photos sometimes look wrong for reflections and surfaces, and how AI shadows work for grounding and light direction.

Does it matter? Yes, more than any other category for overall believability. Viewers feel lighting inconsistency even when they cannot articulate it.

Which artefacts matter for listings — and which do not

Not every artefact is worth your attention. The honest triage:

Always act on: melted badges, degraded wheels, warped panel lines, anything that changes how the car itself reads. These touch accuracy, not just aesthetics — and accuracy is the line that matters, as covered in are AI car photos misleading?.

Act on if visible at full size: halos, reflection mismatch, shadow direction errors. These make the image feel fake without misrepresenting the car.

Usually ignore: background texture repetition, minor softness in areas away from the car. Invisible at the sizes buyers actually view, and invisible in marketplace thumbnails entirely.

A useful calibration: most listing photos are judged in under two seconds at less than half their pixel size. An artefact you need a magnifier to find is not costing you an enquiry. An artefact you can see in the thumbnail is costing you trust on every car around it.

The pre-publish check ritual

The position behind all of this: Motuva does not generate the car. Your car's own pixels are cut from your photograph and composited into a rendered studio, so every artefact above that touches the vehicle — melted badges, lost spokes, bent shut lines — cannot happen to it. Check the output against the original anyway. The cut-out edge is where a difficult shot can still go wrong, and the listing is yours. Make it a ritual, not an intention. Per image, it is about ten seconds:

1. Thumbnail test. View the output small. Does anything look off at a glance? That is what most buyers will see.
2. Car detail pass. Full size: badges, grille, wheels, trim — compare against the original.
3. Line pass. Trace the roofline, shoulder line, and sills with your eye. Straight things straight.
4. Light pass. Shadow under the car, reflections on paint and glass, overall light direction. Does the scene agree with itself?
5. Honesty pass. Does the car in the output match the condition of the car in your original? If anything on the vehicle looks better or different than reality, do not publish it.

Fail any step: with a generative tool, re-run the image — second passes often land clean. With a deterministic one like Motuva, a re-run returns the same picture, so re-shoot the frame or fall back to the original photo for that shot.

If you want the fuller picture of how the output process works, the Motuva FAQ covers it directly. And the cheapest way to calibrate your own eye is to process real stock and inspect the results yourself — the free tier is 20 images a month, full quality, no card required.

Start free — no card, no demo →