A Painted Wall Is Never One Color

How it works

The photograph is scaled to about one and a fifth megapixels and repainted at 0.4 denoise, the lowest setting in this family, because this is the smallest edit in it. The instruction changes exactly one property — the color of painted wall surfaces — and names what must survive: the bright and dark areas across each wall, the shadows furniture casts on them, and the falloff into corners.

The chosen preset is mapped to a fixed color description on the server rather than passed through, so an unrecognized value falls back to the neutral instead of reaching the model.

The negative prompt leads with the failure that defines this page, which is a flat evenly colored wall, and then lists the encroachments: painted trim, painted skirting, painted ceiling, painted furniture.

The shading is the photograph

It is tempting to think of a repaint as a fill operation, the way a paint bucket works in an image editor. That intuition is what produces bad results.

The pixels on a wall encode two things multiplied together: the color of the paint, and the amount of light reaching that spot. A repaint has to replace the first and leave the second exactly alone. Get that wrong and you have replaced the light in the room as well as the paint, which is why flat repaints look like a sticker rather than a wall.

This is also the reason the denoise is held low. A higher setting gives the model more freedom, and more freedom here means it starts reconsidering the light, which is the one thing that was already correct.

The render is not the color

Every visualizer in this category invites the same mistake, so it is worth naming plainly: what you are looking at is not a paint color.

It is a generated impression of a verbal color description, rendered under the light of one photograph taken at one time of day. Three approximations stacked. Any of them can shift the result by more than the difference between two paint chips you were choosing between.

The room's light is the biggest of the three and the least obvious. North-facing rooms lean cool, west-facing rooms swing warm in the evening, and a bulb's color temperature moves everything again after dark. The same tin genuinely looks different in two rooms of the same house, and no render can tell you which of those effects you will be living with.

What this page is good for is elimination. Five colors become two in a minute, and that is a real saving. The final decision needs a tester patch on the actual wall, looked at in the morning and again at night.

Where it breaks first

Two specific failures, both worth checking at full size rather than at a glance.

The first is the edge. Wall meets skirting, casing, ceiling, and switch plate along a line the model has to infer, and inference at a boundary is where a low denoise pass still slips. Follow the skirting along its whole length.

The second is a wall that carries almost no light. A corner in deep shadow has very little color information in it, so what comes back there is largely invention. If your photograph is mostly shadow, most of your render is a guess about how a color you have not seen behaves in the dark.

Publication gate

This page ships once the workflow has been run against a wall with strong window falloff, a wall lit by a single lamp at night, a wall with deep skirting and a door casing, and a wall interrupted by a large piece of furniture, with the luminance profile compared against the source in each case and the skirting line inspected at full resolution.