The photograph decides the result far more than the settings do. Here is what survives being reduced to a grid, and what never does.
A chart is a very low-resolution image: 70 stitches wide is 70 pixels of information, less than an app icon. Anything that needs fine detail disappears, and no setting brings it back. A well-chosen photograph, though, gives something recognisable and often prettier than the original: reduction simplifies, and simplification flatters.
This is the only criterion that really counts. A dog filling three quarters of the picture makes a readable chart; the same dog ten metres away in a garden makes a brown smudge. If you are choosing between two photographs, take the one where the subject is biggest, even if it is less sharp: sharpness will be lost anyway, framing will not. Crop before you convert.
A slightly blurred but well-contrasted photograph makes a good chart. A perfectly sharp one whose subject sits at the same value as its background makes none, because the converter picks its colours by difference and there is none. Think of a black cat on a dark sofa, or a white dress against a white wall. Squint at your photograph: if the subject still separates, that is a good sign.
A shelf, foliage, a street: the converter spends its colours describing that clutter instead of the subject. Three fixes, best first: crop tighter, remove the background (the tool does it, with a model that runs on your own machine), or drop the thread count until the background flattens into a single mass, which is sometimes rather good.
This is the subject that works best, by a distance. Fur reduces well because it is made of areas of close values rather than of detail; eyes and nose supply the dark points that make it a face. Aim for a photograph taken head-on, at the animal's height, outdoors or near a window. A black animal needs more threads than others, not fewer: all the modelling lives in the very dark greys.
This is the hardest subject and it is better to know beforehand. A face reads on tiny differences in skin tone, and the human eye notices immediately when they are wrong. Below about 90 stitches wide, a face becomes someone vaguely similar; it often takes 120 and fifteen threads to get an actual person back. A profile, a backlit silhouette, or a very contrasty black-and-white portrait do far better than a front-lit face in soft light.
A single flower, a bunch, a bicycle, a house: anything with a clear shape and little background works. Landscapes are a special case: a whole panorama has no subject, so it makes no chart, but a lone tree, a boat, a horizon with a lighthouse on it, yes. Again: look for the shape, not the detail.
They can straighten a photograph taken sideways, lift the colour when the result looks washed out beside the original, cut the background away, and swap a thread you dislike. They cannot invent detail that is not in the image, nor separate a subject from a background of the same colour. The best setting is still a better photograph.
It costs nothing, it takes a minute, and the chart is yours.