What reduction actually asks of a photograph
Turning a photograph into a paintable guide is one operation repeated: decide where a region ends. Everything else follows from that. If the boundary between two areas can be found, they become two shapes with two numbers. If it cannot, they become one shape, or worse, forty small shapes strung along a boundary the engine kept changing its mind about.
So the question to ask of a photograph is not whether it is beautiful. It is whether a person could draw an outline around the things in it. If you can imagine tracing it, the engine can find it. If you would have to guess where one thing stops and the next starts, the engine guesses too, and its guesses arrive on your page as shapes you cannot paint.
What reduces well
Four properties, and a photograph needs most of them rather than all of them.
- Separated colour. Areas that differ in value as well as in hue. A red jumper against a green hedge reduces beautifully. A red jumper against a brown fence of the same brightness does not, because to the engine those are one region.
- Edges that stay put. A cat on a plain rug has one boundary and it is decisive. A cat in long grass has a boundary that dissolves into a hundred small ones.
- A subject that occupies real space in the frame. Something filling a third of the picture gives the engine room to give it several regions. Something occupying a twentieth gets one.
- Light with a direction. A single light source produces shadows, and shadows are what turn a flat shape into a form worth painting. Flat overcast light gives correct colours and nothing to paint.
What fights back, and why
These are the four that arrive in support requests most often, and none of them is a fault in the picture. They are pictures that happen to be hard to divide.
A gradient sky is the clearest case. To your eye it reads as one continuous thing. To anything that has to place a boundary it is a smooth ramp with no boundary anywhere, so the engine puts one in wherever the banding falls. You get five stripes across the top of your painting that were never in the photograph and cannot be painted convincingly.
A very dark interior is the second. A camera compresses everything in shadow towards the same few values, so a room that your eye read as full of separate objects arrives as one large dark mass with some furniture rumours in it. The guide is technically correct and describes a picture nobody took.
The third is a busy background. A portrait taken in a kitchen has a face worth six regions and a background worth two hundred, and the detail control cannot tell the difference between them. Lower it and the face goes flat. Raise it and the kettle gets more attention than the person.
The fourth is fine repeating texture. Fur, foliage, gravel, knitwear. Each individual mark is far below the size a brush can paint, so either they are dropped, which reads as plastic, or they are kept, which produces a page of confetti.
None of these is a reason to give up on a photograph. They are reasons to crop it, or to change what you are asking for, which the last section covers.
Two things that surprise people
Sharpness matters far less than most people expect. A slightly soft photograph reduces about as well as a crisp one, because the operation is finding regions rather than resolving detail. A blurred photograph with clear masses will give you a better guide than a razor sharp one of a cluttered scene.
Resolution matters less than expected too, within reason. The engine works at a fixed internal size, so anything beyond about 1200 pixels on the long edge is discarded before analysis begins. A 3 megapixel phone photograph and a 45 megapixel camera file of the same scene produce very nearly the same guide. What the larger file buys you is the freedom to crop hard and still have enough left.
What to do with a photograph that will not reduce
Crop first, before touching any setting. Most difficult photographs are difficult because they contain two subjects competing for the same detail budget. Removing the kitchen from behind the portrait fixes more than any control on the page.
Then change what you are asking for rather than how much of it. A picture that produces an unusable guide at 32 colours often produces an excellent one at 8, because at 8 the engine stops trying to describe the gradient and starts describing the shapes underneath it. Reducing further is frequently the answer, and it is the opposite of what people try.
If the picture still refuses, ask for a different kind of structure. Shape Colouring lays a repeating tiling over the photograph instead of tracing it, so a gradient sky becomes an orderly field of hexagons that steps through the gradient rather than a set of accidental stripes. Foliage and fur become texture rather than confetti. It is the answer for exactly the pictures that defeat contour tracing.
And some photographs are not going to work. A dim group shot at a party has no separated colour, no directional light, no dominant subject and a background of other people. There is no setting behind which a good guide is hiding, and the honest thing is to pick another picture.