Lucasgitguy-co/forma3d-tiny-2

Forma3D Tiny 2

A 306,321-parameter image-to-Gaussian research model continued from Lucasgitguy-co/forma3d-tiny at revision e80b8d895cc2e405894473abe4ce40b3472eaca7. The optimizer starts fresh. It uses a 16 cubed grid and width 64. Selected checkpoint step 7800 from 8000 completed steps.

Training

See training.json and metrics.jsonl for actual optimizer settings and validation history. Inputs are orthographic renders of centered mesh-derived Gaussian proxy clouds on black. Targets are isotropic surface samples compressed to one Gaussian per occupied grid cell. No pretrained 3D teacher was run. No natural-photo training or differentiable rendering is claimed. Asset IDs, known source groups, and exact mesh hashes from all 96 known v1 assets were excluded. Near duplicates can remain. The separate 15 percent object-group test split does not enter checkpoint selection; 20 percent of development groups supply validation. Counts are realized counts in preparation.json and data-audit.json, not assumed independent asset counts.

Evaluation

Both checkpoints were evaluated on the same new held-out proxy data.

Metric	Starting Forma3D Tiny	Selected Forma3D Tiny 2
Mean test loss	0.616513	0.570490
Mean view occupancy IoU at threshold 0.5	0.286327	0.327870

Test groups: 36; views: 156. See evaluation.json for hashes and release gates. Metrics are view-averaged over proxy renders. They do not establish photo realism, novel-view image quality, mesh quality, texture fidelity, gameplay, or general-purpose 3D reconstruction.

Use

Install this source with Python 3.11 and uv pip install ".[data]" in a virtual environment, then run:

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

Output is Gaussian PLY. Inputs should match centered objects on black. See COMMANDS.md for the complete continuation recipe. This model retains the same architecture; no text-prompt conditioning or new mesh decoder was added.

License

Publisher source and weights are offered under Apache 2.0. Retain NOTICE, THIRD_PARTY.md and ATTRIBUTIONS.json. CC BY Scenario assets and CC0 Objaverse assets have their own attribution and license records in DATASETS.md.