PhotoTo3D DOCS
Support

Troubleshooting

Common reconstruction errors and how to get a better model.

Reconstruction returns an error

A 503 means reconstruction is not configured on the server — the OPENAI_API_KEY Worker secret is missing. Set it before the endpoint can call the vision model.

A 413 means the processed reference is too large. Send a smaller image; Studio already resizes before upload, so this usually points at an oversized direct API call. A 422 is a refusal from the vision model — it declined the specific image rather than failing technically. Rephrase the intended use or choose a different reference.

A 401 or 402 is about access, not the image: 401 is a missing or unrecognized API key, and 402 means the account's monthly reconstruct credits are exhausted.

The model looks wrong

If proportions or hidden parts look off, the reference is usually the cause. One image cannot show the rear face or internal construction, so the system infers those regions and marks them in the spec's uncertainties. A flat, straight-on photo gives it the least depth to work with.

A poor camera angle produces confident but wrong geometry. A three-quarter view that reveals depth, side profile and attachment points lets the reconstruction place parts correctly instead of guessing.

How to get a better result

Use one clear image with the whole subject visible, filling more than half the frame, without motion blur or heavy occlusion — the same guidance as the Known limits guide. Sharper evidence yields fewer inferred regions.

For a production hero asset, provide consistent front, side and rear views so less has to be inferred. And match the method to the subject: hard-surface props and product-like objects reconstruct well, while dense organic forms and photoreal faces are better served by mesh or neural reconstruction.