Forma3D-Tiny v0.2

An experimental compact image-to-Gaussian model post-trained from Image3D Student's synthetic demonstration checkpoint. This is a coarse research baseline, not a general-purpose photo-to-3D model. Its grid is 16 cubed and it has 306,321 parameters. Selected checkpoint: step 4800; completed run: 5000 steps.

Actual training data and preprocessing

Only the datasets listed above were used, with per-asset provenance in ATTRIBUTIONS.json. Scenario assets are CC-BY-4.0; the Objaverse subset is CC0. Original objects were centered and uniformly scaled into a 0.9-wide cube, surface sampled, and converted to isotropic Gaussian proxy targets. Orthographic images were rendered from these same proxy clouds. This is not teacher distillation from Idol, Hunyuan, InstantMesh, or Hi3D, and it did not train on natural photographs, MultiScan, 3D-Front, Show3D, or Voxel51 scene splats.

Related source inputs and all views stay in the same object group. A separate 15% test partition is excluded from training and validation-based checkpoint selection.

Measured results

Independent test objects: 14; test views: 60. Base test loss: 0.842823; post-trained test loss: 0.584055. Base occupancy IoU: 0.272567; post-trained occupancy IoU: 0.307005. See evaluation.json for the full results and checks. These scores measure coarse proxy reconstruction on the rendered data distribution; they do not prove photo realism, watertight meshes, production quality, or texture fidelity. Near duplicates can remain despite source grouping and exact-byte deduplication.

Run on a Mac

Download this repository, then run bash setup-mac.sh and source "$HOME/Library/Application Support/Forma3D/venv/bin/activate".

image3d generate --checkpoint model.pt --image your-centered-cutout.png \
  --out results/object.ply --device mps
image3d preview --ply results/object.ply --out results/preview.png

Inputs are centered objects on black, matching the training framing. Natural photos are outside the verified distribution. Output is Gaussian PLY, not a triangle mesh. CPU works with --device cpu; CUDA support is implemented but was not tested on the Mac used to prepare these scripts.

License and attribution

Source code and these weights are offered under Apache-2.0. Dataset and individual asset licenses remain separate; retain ATTRIBUTIONS.json and consult DATASETS.md. This license statement does not grant rights to third-party trademarks or likenesses. See COMMANDS.md for reproduction commands.