الخلاصة (BLUF): Upload a photo of a single object — TripoSR reconstructs the volume and returns a GLB file. Everything runs in your browser on your device GPU: the photo is not uploaded, no tokens are charged and no internet is needed after the model is cached.
The TripoSR model (MIT, Stability AI + Tripo AI) is about 485 MB, downloaded once into the browser cache and afterwards works offline. The input is a 512×512 photo; reconstruction runs on your device GPU via WebGPU.
The flow: pick a photo, the background is removed automatically, and the model reconstructs the volume from a single view and fills in the hidden sides; the grid resolution is adjustable (128 is faster, 256 is the default).
One object, a plain background, a front view in good light. A transparent PNG gives the cleanest model: the cutout reads the alpha channel directly; on flat photographic backgrounds a flood cutout from the borders does the job.
The model handles a single object per photo: two objects in the frame merge into one shape; a middling result is not a failure but a property of single-image reconstruction.
Completely free: no tokens are charged and there is no paid API. The photo stays in your browser and is never uploaded; the model runs locally.
Yes. The model runs on your device GPU via WebGPU — a recent Chrome or Edge (desktop or Android). A clear message appears on devices without WebGPU.
No. The photo stays in the browser and the model runs on your device; nothing is uploaded to the server.
About 485 MB, downloaded once into the browser cache — afterwards generation works offline.
Try a photo of one object on a plain background, front view, no overlaps — TripoSR quality depends heavily on the angle and light. No tokens are charged for attempts.