TripoSR fp16 ONNX for browser WebGPU

Single-image to 3D reconstruction. TripoSR exported to ONNX (fp16) and split into three sessions so it runs entirely in the browser on onnxruntime-web with the WebGPU execution provider (GridSample runs natively).

Files:

encoder_fp16.onnx (~173 MB) - ViT image encoder. Input image (1,3,512,512) fp32 -> tokens (1,768,1025).
backbone_fp16.onnx (~666 MB) - triplane transformer. Input image_tokens (1,1025,768) -> scene_codes (1,3,40,64,64).
decoder_fp16.onnx (~0.1 MB) - triplane MLP. Inputs triplane (3,40,64,64) + points (N,3) -> density (N,1), color (N,3).

Model I/O is fp32; weights are fp16. Preprocess: RGBA composited over 0.5 gray, bilinear antialias resize to 512x512. Decode a grid over (-0.87, +0.87), marching cubes at iso 25 on the density field.

Credit

Original model: VAST-AI-Research/TripoSR (Tripo AI + Stability AI). Apache-2.0. This is a repackaging of those weights to ONNX for in-browser inference.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Image-to-3D

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support