The best depth map generator in 2026 is the one that shows its work

Search results for depth map generators mix three jobs. A free demo returns a preview image and stops. A production tool returns files that another program can consume: raw depth, 16-bit maps, meshes, and a manifest. An API returns a depth image to a server and leaves storage, scale, and manufacturing to the caller. A tool can do more than one, but it cannot be judged on a single screenshot.

The distinction matters because the failure costs are different. A weak preview wastes a few seconds. A weak production file wastes a material blank, a machine slot, or a customer deadline. Before comparing outputs, decide which job you are buying. DepthCarve is built around the second job and exposes the first as the free on-ramp.

First, does the tool inspect the source? Contrast, edge detail, border clutter, and resolution are cheap to measure and expensive to ignore. Second, can you compare the estimate with the original instead of judging a grayscale image in isolation? Third, can you change the useful range without triggering another paid model call? Fourth, can you export a test file that opens in the program that will actually make the part?

Features that do not answer one of those questions are decoration. A gallery of pretty examples does not tell you whether your photograph has a usable tonal range. A long list of formats does not tell you whether the 16-bit file contains 16-bit information. The preview earns trust by exposing the decisions, not by hiding them behind a Generate button.

A credible bundle starts with raw floating-point depth. That is the source the other files derive from. The two 16-bit PNGs should be verified as 16-bit and labelled as forward and inverse maps, not as two copies of an 8-bit preview. The mesh should be closed, with a top surface, side walls, and a bottom face. A manifest should record the source fingerprint, adjustments, dimensions, machine assumptions, and validation results.

This is also where many tools become vague. Ask whether the 16-bit file was encoded from a float field or upconverted from an 8-bit preview. Ask whether the STL is watertight or an open height sheet. Ask whether a failed provider call returns the credit, and whether a retry can debit twice. If the answers are not visible before payment, the price is not the only unknown.

Depth estimation is probabilistic. A provider can time out, a queue can stall, and an output can arrive without the contracted raw depth. A tool that treats every failure as the customer's problem is more expensive than its price. The minimum standard is an idempotent debit and an automatic refund when the provider fails or cancels.

There is a second, subtler boundary: a completed estimate that is simply unsuitable for the material. That is not a technical failure, and a tool should say so plainly. The free preview exists so that decision happens before a credit is spent. A refund policy that covers provider failure but not taste is honest; a policy that promises a perfect physical result from one photograph is not.

The product started from a simple observation: the model call is the shortest part of the job. The expensive part is discovering that a background carries more depth variation than the subject, or that a hard shadow became a ridge, after the material is already on the machine. We published two preview cases with measured evidence rather than a gallery of untraceable results. In the busy-background case, the outer border band spans 99.2% of the depth range and has a standard deviation of 73.5 against 55.3 in the subject. In the strong-shadow case, source brightness and returned depth correlate at r = 0.580, which is exactly the confusion a single-view model cannot fully resolve.

Those numbers are not marketing claims. They are the reasons the workbench shows source checks, a comparison slider, and manufacturing limits before the approval button. They also explain why we do not sell an unlimited plan or claim metric truth from one photo. The honest product is a review loop with a bounded paid output.

Open the candidate with a portrait, a product photo, and a deliberately cluttered workshop image. Check whether the source report changes in a way that matches what you see. Move the black and white points and confirm the 3D preview updates without a new inference charge. Download the free PNG and preview STL, then open both in your real software. Finally, read the pricing and refund terms and ask what happens when the provider fails.

If the tool cannot pass those steps, a better model will not save the workflow. If it can, the remaining question is whether the production bundle contains the files your process actually consumes. That is a narrower, more useful comparison than a leaderboard of grayscale thumbnails.