Forma3D Tiny 3

Continued from Lucasgitguy-co/forma3d-tiny-2 at 1648582b07a0f7f9d482a8c0f4cf89e6ecd86972. Same 306,321-parameter architecture, 16 cubed grid, width 64, maximum 4,096 Gaussian slots. Fresh AdamW optimizer at the start of continuation. Completed 6000 steps; selected step 6000 by validation loss only.

Data and scope

512 ABO objects, 2048 orthographic Gaussian-proxy images on black. Data revision: c7c8293fed3b7ab122ef5d7896521a187e0962d7. Targets are isotropic mesh surface samples, compressed to one Gaussian per occupied grid cell. Texture surface colors are sampled; full PBR rendering and differentiable rendering are not used. No natural-photo training is claimed. MeshAnythingV2, Idol Diffusers, SHARP, Pixal3D, Idol Turbo and Qwen3.5-4B-3D-Coder were not run, merged, or distilled in this Mac-only recipe.

Objects, shared Amazon item IDs and exact mesh/image/target hashes are grouped before splitting. Split groups: {"train": 348, "validation": 87, "test": 77}. Test data is excluded from training and checkpoint selection. Near duplicates can remain; overlap with ancestor training datasets is not fully known.

Evaluation

Both checkpoints were evaluated on the identical held-out ABO proxy views.

Metric	Starting v2	Selected v3
Mean test loss	0.690551	0.544487
Mean view occupancy IoU, threshold 0.5	0.286866	0.339860

These measurements do not establish natural-photo reconstruction, photorealism, mesh quality, novel-view rendering quality or improvement on v2's original domains. No confidence intervals or broad generalization claim. See evaluation.json, preparation.json and split manifests.

Use

With Python 3.11, install this repository using pip install ., then:

image3d generate --checkpoint model.pt --image centered-object-on-black.png --out object.ply --device mps
image3d preview --ply object.ply --out preview.png

Output is Gaussian PLY, not a textured GLB mesh. Inputs should match centered objects on black. The model has no text prompt conditioning. Reproduction launcher and locked dependencies are in recipe/.

Attribution

Publisher source and weights: Apache 2.0. Retain inherited NOTICE, THIRD_PARTY.md and ATTRIBUTIONS.json. ABO models, images and listing text: Amazon.com and the Amazon Berkeley Objects contributors, CC BY 4.0; converted to GLB by CodePawl. The recipe further decompresses geometry, normalizes coordinates, samples colored surfaces and renders proxy images. See ABO dataset, ATTRIBUTIONS.json and licenses/LICENSE-ABO-CC-BY-4.0.txt. Raw assets are not bundled.