A reliable workflow
Pick the channel that carries structure
Perceptual luminance weights green most strongly; channel modes are useful when one channel isolates material structure better than the composite.
Remove broad lighting before estimating relief
Broad illumination should usually be removed before fine relief is extracted, otherwise the virtual surface leans toward the photographed light.
Blur noise without erasing the material
Denoising before gradients suppresses pixel noise, but too much blur removes real pores, grain, and bevel transitions.
Decide what brightness means before estimating depth
An image to normal map conversion is an interpretation, not a depth measurement. First identify whether luminance, a single color channel, or alpha best follows the surface structure. Paint, stains, reflections, and cast shadows can be mistaken for height, so remove broad lighting gradients and mask unrelated color where possible. Compare the generated image to normal map result with a known raised or recessed feature. If the source cannot support a consistent high-versus-low reading, create a dedicated height map instead of increasing strength until the preview looks dramatic.
Control noise without polishing away the material
Fine JPEG blocks and camera noise create alternating gradients that become sparkle in an image to normal map result. Add only enough pre-blur to suppress detail smaller than the intended texel footprint, then inspect both full resolution and a reduced view. Scharr is a sound general choice for balanced diagonal response, while Sobel may be useful when matching an established pipeline. Export lossless PNG and verify the image to normal map output in the destination renderer; the purple thumbnail cannot reveal an incorrect Y convention, color-space transform, or tangent mismatch.