stabilityai/TripoSR

Methodology

• Latency: warm measurements: two warm-up requests, then ten timed requests per evaluation image; stage timings include host/device copies and synchronization. GPU-only timings exclude copies.
• Accuracy: The reference and candidate use the same crop, density threshold 25 and 128³ grid; the PyTorch reference resizes with PyTorch and extracts with scikit-image, the served candidate with Pillow and scikit-image. Chamfer is the average of the two directional mean Euclidean distances, divided by the reference bounding-box diagonal. F-score uses a 1% diagonal threshold and 100,000 sampled surface points per mesh. No independent alignment or scaling is applied. Calibration uses different camera views from evaluation; four synthetic objects measure precision agreement, not real-world reconstruction quality.
• Build: This Embedl build serves three native TensorRT engines with FP16 encoder, backbone and density/colour MLP compute; sampling and output activations remain FP32. No PyTorch is present in serving. An Embedl Deploy INT8 encoder failed the same four-object mesh gate on H200: chair F-score 0.889033 and table 0.978028 (required >=0.98), versus 1.0000 for FP16 on every object. Ship FP16; no INT8 variant is published. These synthetic views measure numerical agreement, not general reconstruction accuracy.
• Serving: Serving uses no PyTorch: three TensorRT engines (the DINO image encoder; the FP16 triplane tokenizer, backbone and upsampler; the FP16 density/colour MLP with FP32 sampling and activations) and NumPy, in the shared TensorRT runtime. Marching cubes is scikit-image (BSD-3-Clause), shipped in the model under lib/. Stage timings include host/device copies and synchronization: grid reuses cached coordinates and includes all eight query batches; colour is the vertex query; encode includes GLB, PNG and base64. Warm measurements use two warmup requests then ten timed requests per evaluation image. GPU-only timings use populated buffers after a real request, excluding copies. INT8 here describes only the encoder. Static examples link their GLB and PNG crop from the public embedl/documentation-images Hugging Face dataset; live responses contain base64 data. The input chair is original procedural geometry rendered by Embedl’s synthetic.py, dedicated to CC0-1.0; no external photograph was used. It is also included under examples/.
• Baseline PyTorch: the earlier serving of the same FP16 model: a TensorRT image encoder with the triplane generator and the density/colour MLP in PyTorch, same device, same measurement.