Trace a PNG into editable vector paths

No, and the page will tell you so with a number rather than letting you find out later. Tracing needs a boundary, and a photograph has none: the pixel values change continuously, so reducing them to 16 fills produces the posterised bands you get from a screen-print filter, cut into thousands of rings. The mean ΔE readout above the output measures exactly this — under 2 the palette fits the artwork, over 6 you are looking at shading, and in the teens you have a photograph, at which point the honest advice is to keep the raster and export it at the size you need instead.

No, and that distinction is the reason this page exists. A large share of converters that answer to this search wrap the PNG in a base64 data URI inside an image element and rename the result .svg: the file has the extension and none of the properties, since it will not scale cleanly, cannot be edited, and lands roughly a third larger than the PNG it swallowed. Any converter takes five seconds to check — open the file in a text editor and look for the string base64. Here you will find nothing but M, L, C and Z followed by coordinates.

Because the antialiased pixels along an edge have to be assigned to one side or the other before anything can be traced. A blend pixel that is 60% blue and 40% white is quantised to whichever palette color is nearer in Lab, and the boundary therefore lands on the pixel grid rather than between the grid lines. Three things help: trace on the 1024 px grid instead of 512 so each edge decision is finer, switch corners to Curved, which fits a Catmull-Rom curve through the simplified nodes, and raise the tolerance to about 1.2 so the staircase is replaced by a straight run rather than followed faithfully.

Yes — each color arrives as one compound path with real anchor points, so a node tool grabs them and a fill swatch recolors the whole layer at once. The one thing to know before you start pulling nodes about is that every path carries fill-rule="evenodd", which is how the counters survive: the inside of an O is a second ring within the same path, and the even-odd rule is what punches it out. If your editor rewrites the fill rule to nonzero on import, holes will fill in, and the cure is to reverse the direction of the inner ring rather than to redraw it.

It comes out as outlines, not as text, because nothing here reads letters. There is no OCR step: a capital R is a shape with a hole in it as far as the tracer is concerned, and the SVG will contain path data rather than a text element with a font-family. For a wordmark that is usually the correct outcome, since outlines are exactly what a logo file should hold. For a paragraph it is not — retype it in your design tool, where it stays searchable, selectable and a tenth of the size.

Because vector data is only smaller when the drawing is simple, and the byte counts on screen let you see where the crossover sits for your image. The three-color mark used to measure this page came to 7.4 kB of path data against the 24.8 kB PNG it was traced from, and the gap widens in the vector's favour every time the display size goes up, because the path data does not change. A busy or noisy image goes the other way fast: raise the tolerance, cut the color count, or turn the island slider up, and watch the node total fall. If none of those get the file below the PNG, that is the measurement telling you this picture wants to stay a bitmap.

Two controls throw detail away deliberately, and both are yours to switch off. The island slider merges any connected area smaller than its value into the neighbor it shares the most edge with, which is what removes JPEG speckle and stray anti-alias crumbs — set it to 0 and every speck is kept. Detail tolerance discards nodes that sit within its distance of the straight line between their neighbors, so a one-pixel notch disappears at 1.5 and survives at 0. Trace at 1024 px with both at zero if you need everything, then simplify from there.