Hi3Dv1.0-mlx · W8A8

← Project & inference runtime · Results

Experimental; M5 family required. Signed-H64 INT8 transformer projections with dynamic activation quantization and native integer GEMMs. Distinct from affine INT8 weights; requires the included W8A8 runtime. Quality and speed vary.

Selective download	Transformer(s) + shared VAE
512³	2.38 GiB
1024³ / 1536³	4.41 GiB

These are checkpoint sizes, not runtime memory. Encoders and Python dependencies are additional. The shared components are stored in Hi3Dv1.0-mlx-shared.

Run it

Follow the main installation guide, then:

hi3d-mlx refine --image reference.png --coarse-mesh coarse.glb \
  --preset w8a8 --resolution 512 --output refined.glb

The CLI downloads only this variant, the shared VAE/configs and the pinned image encoders. No conversion step is required. For an offline setup, run hi3d-mlx download --variant w8a8 --resolution 512 once, then add --local-files-only to inference. Generic mlx-lm loaders do not implement this geometry pipeline.

Evidence and limitations

See the benchmark methodology and complete recorded results. The historical cohort used Apple M5, 32 GiB, MLX 0.32.4. Results are geometry only and require both an image and an existing coarse mesh. The 1536³ grid reuses transformer_1024; it is not a fixed face count.

Format and attribution

Experimental signed-H64 native INT8 weights and activations; requires the compatible W8A8 runtime path.

The transformer weights in this repository were modified by Devin Lai through MLX-format quantization. The conversion implementation is private; the inference implementation is published in the main project. All applicable upstream notices and the Hi3D Open License Agreement v1.0 are retained. No blanket MIT or Apache license is asserted for the model.

All index paths are relative to the assembled model root. shared_components are resolved from the shared repository, not this transformer repository.