xjodoin/moge-3-vitl-webgpu

MoGe-3 ViT-L WebGPU

Unofficial FP32 ONNX and packed WGSL weight exports of Microsoft's MoGe-3 ViT-L, including the learned sparse 3D refiner. The original ONNX/WGSL files and lossless packages preserve FP32 weights. The optional progressive package rounds selected encoder weights for smaller downloads and includes correction data that restores every original byte. There is no retraining. These files require the browser-side sparse orchestration in the source project; they are not a single end-to-end ONNX graph.

• Source, setup and complete validation
• Browser demo
• Original model

Profiles: profiles/fast (512 x 344, 1200 tokens) and profiles/detail (768 x 512, 1733 tokens), both supporting one, two or three refinement passes. ONNX comparison blocks are in refiner/; dedicated kernel parameters and checksum manifests are in refiner-wgsl/. Each profile manifest records dimensions, upstream revisions, file size and SHA-256.

All 54 browser-versus-original CUDA FP32 comparisons passed on an Apple M4 Pro / Chrome 156, covering three images, two resolutions, one to three passes and three repeats. Maximum point-map NRMSE was 0.0943%, with matching masks. Pointwise outliers are larger. Earlier MLX comparisons remain diagnostic; original CUDA is the native reference. These synthetic-photo parity checks are not physical measurement validation. See the included benchmarks and the source README for exact gates, timing, memory accounting and limitations.

Requires approximately 1.5 GB of weights per selected profile and several GB of available memory. The browser uses WebGPU for learned operations and voxel indexing, with JavaScript orchestration and camera geometry. No photo upload or server inference is necessary.

When included, packed-v2/ provides adaptive lossless codecs and indexed bundles: about 1.24 GB and 89 model requests for a cold profile, 6.00 MB extra for Detail after cached Fast, and zero model downloads on cached revisits. Older v1 cache entries remain reusable. The original packed/ layout is retained for compatibility. All original FP32 model bytes are preserved. See ADAPTIVE.md for validation, setup and storage limits.

When included, packed-v3/ provides selective 24-bit encoder storage with FP32 computation. Compact weights are approximate. The original-weight mode reuses the compact download and fetches separate correction data to reconstruct original FP32 bytes, verified by their original hashes. Existing full-precision cached weights can serve both modes without downloading replacement weights. See PROGRESSIVE.md for measured download sizes and the separate compact/exact accuracy matrices.

Pinned original checkpoint: 184008f877d7ad1ad4c2cd2182a9bd1f63d0e5be; MoGe source: 74fbce054ebed49800de42d0ad0e83495065719a. Upstream code, weight and dependency licenses are included in licenses/. This is not an official Microsoft release.