JPG to STL: fix the source before the mesh

JPEG divides the image into blocks and discards high-frequency information that the eye is less likely to notice. Around hard edges, that shows up as ringing and blocky halos. A depth model reads those brightness changes as shape. The result is a relief with small terraces along the silhouette, a rough background, or a halo that looks like a raised outline around the subject.

The effect gets worse when the JPEG was saved several times, heavily sharpened, or resized after compression. A screenshot of a JPEG can add another resampling step. If the original PNG or camera raw is available, use it. If not, inspect the JPG at 100% and crop away the worst compression zones before generating.

A photograph records light, not distance. A bright forehead and a bright wall can share a value while sitting at different depths. A hard shadow is dark and sits directly against the object that cast it, which is also the pattern that signals proximity. The model can turn that shadow into a ridge because the single view does not contain enough information to separate the two.

Soft, even lighting describes form more reliably than direct sun or a single hard flash. Backlighting, lens flare, and strong specular highlights are also unreliable. If the only available photo has a hard shadow across the face, crop it out, flatten the background, or choose a different photo before spending a credit.

Use the largest original, not a messaging-app copy. Check that the subject is sharp at 100% and that the eyes, product edges, or architectural lines are not smeared. Remove captions, stickers, and UI overlays. Keep a quiet border if the result will become a plaque or tile. Avoid filters that add vignettes, clarity, or HDR halos.

If the image is a scan or a screenshot of a document, the depth interpretation is usually meaningless. Use an extrusion or vector workflow for logos and line art instead. If the image is a photo of a photo, watch for glare and a visible frame; both can become geometry.

The first free preview is a hypothesis. Compare it with the source: are the foreground and background in the right order, are the edges where they should be, did the shadow become a ridge, did the background flatten? If the answer is no, do not proceed to a production mesh. Fix the source or the range first.

The 3D preview is intentionally low resolution. Use it to check orientation, scale, and gross shape, not to judge fine detail. The free preview STL lets you open the geometry in a slicer or CAM package before paying. If the preview already shows a broken border or a noisy background, a denser mesh will preserve the same problem.

If you already have a height map, use PNG rather than JPG. Height maps are data and should not pass through a lossy codec. If the source is an 8-bit grayscale image, keep it as PNG and do not upconvert it to 16-bit and call it precision. If you need the true 16-bit pair, approve a production job so it can be encoded from raw float depth.

For photographs, the camera format matters less than the capture quality. A clean, sharp, well-lit JPG can outperform a noisy PNG. The question is not the extension; it is whether the pixels still describe the form you want to manufacture.

Set the finished width, relief height, base, and mesh grid before approval. A JPG converted at 50 mm and the same JPG at 300 mm need different relief heights and base thicknesses. Check the bounding box, manifold status, and the thinnest feature in the target program. For printing, slice a small test; for CNC, simulate the toolpath and cut scrap.

Keep the original JPG, the approved ZIP, and the manifest. If a customer asks for a larger version, the recorded dimensions and adjustments are the starting point, not a guess.