Forma3D Tiny 4

A 306,321-parameter image-to-Gaussian research model continued from Forma3D Tiny 3 at 3469f527c5f6833d654f62653ee92e0f73854b33. The architecture remains a 16³ grid, width 64 and at most 4,096 Gaussian slots. Trained locally on an M4 MacBook Air with 16 GB unified memory using PyTorch MPS, float32 and batch size 8.

Use

Clone and install with Python 3.12. Git LFS is needed to fetch the checkpoint:

git lfs install
git clone https://huggingface.co/Lucasgitguy-co/forma3d-tiny-4
cd forma3d-tiny-4
python3.12 -m venv .venv
source .venv/bin/activate
pip install .
image3d generate --checkpoint model.pt --image examples/abo-B071FJR479-v3-input.png --out object.ply --device mps --threshold 0.6
image3d preview --ply object.ply --out preview.png

Use --device cpu on systems without Metal. Replace the sample input with an actual image path to try other inputs. The recommended threshold is 0.6; the inherited CLI defaults to 0.5, so pass the flag explicitly. generation_config.json is informational and is not automatically read by the CLI. Output is a Gaussian PLY, not a textured mesh. No text prompting is supported.

Training and selection

The expanded dataset contains 1,024 ABO objects across all splits, not 1,024 training objects. Training uses 677 objects with 12 views each (8,124 inputs): four azimuths at three elevations. The original 512-object split assignments are preserved; duplicate identities are grouped before new objects are assigned. The common validation set has 87 objects / 348 views. An additional 100 objects / 400 views are reserved for final confirmation. The historical test set is excluded from tuning.

Three variants start from the same v3 checkpoint: original data, 12-view data, and expanded-object data. Each receives 6,000 AdamW steps at learning rate 5e-5. Selection maximizes validation silhouette IoU subject to nonworsening occupancy IoU and Chamfer distance, with thresholds 0.3, 0.4, 0.5 and 0.6 tuned on validation. The expanded-object variant won. model.pt is seed 44, step 6000; repeat/model.pt is its independently trained seed-45 confirmation. The initial winner was fixed before inspecting final confirmation results.

Held-out confirmation results

All rows use the same 100 held-out objects / 400 proxy views. Higher IoU and lower Chamfer distance are better. The first two rows isolate the checkpoint change at threshold 0.5; the third includes the validation-selected threshold change.

Checkpoint	Threshold	Silhouette IoU	Occupancy IoU	Chamfer distance
Original v3	0.5	0.463688	0.337398	0.069585
Tiny 4, seed 44	0.5	0.512971	0.424836	0.053612
Tiny 4, seed 44 (recommended)	0.6	0.540550	0.434657	0.052380
Tiny 4, seed 45 (repeat)	0.6	0.546573	0.412526	0.054105

For recommended seed 44 versus original v3 at 0.5, paired object-cluster bootstrap 95% intervals (2,000 resamples) for candidate-minus-baseline changes are: silhouette IoU +0.07686 [0.06621, 0.08926], occupancy IoU +0.09726 [0.07289, 0.12272], and Chamfer −0.01721 [−0.03762, −0.00622]. These intervals include the threshold change. Second-seed intervals also exclude zero in the favorable direction for all three metrics. Detailed scores and per-example results are in evaluation/ and evaluation.json; splits and source metadata are in splits/.

The 15-minute batch-8 training benchmark measured 650.02 samples/s; this excludes preparation and validation overhead. It recorded about 496 MiB peak process RSS and 141 MiB MPS driver allocation; these overlap in unified memory and are not a measurement of total system memory. CPU/MPS loss and gradient parity passed.

Examples and limitations

Each turntable shows the predicted Gaussian preview on the left and a reference mesh silhouette on the right. The six examples include the two lowest, two median and two highest silhouette-IoU views in the confirmation set; they are not random samples. The reference image is a silhouette, not a photorealistic rendering.

Inputs are orthographic Gaussian-proxy renders on black. Natural photographs, photorealistic rendering, arbitrary viewpoints, thin structures and hidden surfaces remain unvalidated. Centering an ordinary photo does not make it an in-domain input. Chamfer measures symmetric mean Euclidean distance between predicted Gaussian centers and mesh-surface samples in normalized coordinates, not reconstructed mesh quality. Silhouettes compare alpha>=0.5 splat rasterization with triangle rasterization at 96 pixels from four fixed unseen viewpoints. Near duplicates may remain; ancestor training exposure is not fully known. These measurements do not establish broad real-world reconstruction quality.

Reproduction and attribution

recipe/ contains the training/evaluation implementation and dependency lock. Original source and weight lineage is Apache 2.0; retain LICENSE, NOTICE and THIRD_PARTY.md. ABO assets and derived example imagery are credited to Amazon.com and Amazon Berkeley Objects contributors under CC BY 4.0, with GLB conversion by CodePawl. Geometry was normalized, sampled and rendered into proxy inputs and silhouettes. See ATTRIBUTIONS.json, attribution/INHERITED.json, licenses/LICENSE-ABO-CC-BY-4.0.txt and the dataset. Raw GLB meshes and the full training image cache are not bundled.