Fire3D is released only after the public commands pass inference, camera sampling, and rendering checks on each supported dataset family. Generated artifacts and model weights are intentionally excluded from Git.
| Dataset | Validation scenes | Protocol | Inference | Rendering |
|---|---|---|---|---|
| iTHOR | iTHOR_FloorPlan24_physics, iTHOR_FloorPlan306_physics, iTHOR_FloorPlan312_physics, iTHOR_FloorPlan328_physics |
fire3d_video_v1 |
Passed | Passed, 16 views |
| Imaginarium | bedroom_01, bedroom_17, bedroom_35, computer_room_03 |
fire3d_video_v1 |
Passed | Passed, 16 views |
| ScanNet++ | 09bced689e |
fire3d_scannetpp_v1 |
Passed | Passed |
| Single image | 003025, 003084 |
fire3d_single_image_v1 |
Passed | Passed at the input camera |
The iTHOR, Imaginarium, and ScanNet++ protocols enable the background room-box
prior, unit-box pruning, and room_box_isotropic_enclose canonical transform.
The scene audit written beside each reconstruction must report
background_room_box_prior.canonical_transform.applied: true and zero points
outside the canonical unit box. Single-image inference keeps its independent
perception-OBB background transform because a room cannot be estimated reliably
from every input photograph.
fire3d --helpand every public subcommand help page run without imports failing.ruff check .andpytest -qpass in the release workspace.- Every validation scene completes through the public
fire3d infercommand. - Camera sampling and rendering complete through the public interfaces.
- Rendered images are nonblank and receive a visual geometry/alignment audit.
- A clean environment installation and clean checkout reproduce the commands.
- Hugging Face archives pass checksum, whitelist, extraction, and smoke tests.
- Model bundle paths use stable aliases and do not expose private iteration identifiers in filenames, protocols, manifests, or model-card text.
- The sparse-structure VAE encoder and decoder load strictly as a pair, run a
finite
8^3 -> 2^3 -> 8^3forward pass, and complete a one-step joint encoder/decoder training smoke test on one GPU.
The authoritative validation is performed on sl-gpu-02; the local workspace
is a mirror of the committed release state.
The advertised ScanNet++ scene, 09bced689e, was rerun on September 8, 2026
through the public fire3d infer command with fire3d_scannetpp_v1. The run
completed perception, batched reconstruction, camera sampling, and rendering
for all 30 predicted instances. Its background audit records
canonical_transform.applied: true, reduces 63,674 previously out-of-box
points to zero, and bounds the retained 369,344 background points within
[-0.49, 0.49]. The authoritative generated output is
results/release_validation_scannetpp_09bced689e_20260908/ on sl-gpu-02;
it remains excluded from Git.
The public artifacts were verified on September 8, 2026:
- The model repository contains a root
config.jsonquery file, stable model aliases, both shape and PBR HC-VAE encoders, and the sparse-structure VAE encoder/decoder pair. The model manifest records the exact public-code commit and SHA-256 digest of every model file. - All 376 dataset archives match the byte sizes recorded in
manifest.json; the remote manifest is byte-for-byte identical to the staged manifest. - The dataset inventory contains exactly 67 iTHOR, 120 Imaginarium, 165 ScanNet++, and 20 single-image scenes selected by the release whitelists.
- Every published iTHOR and Imaginarium scene includes 60 exact-camera RGB
rerenders. All 11,220 archived images across 187 scenes match the
authoritative
renders_updatedimages byte-for-byte. - The optional ShapeR evaluation archive contains 178 pickle-free NPZ samples and includes its source attribution and license.
- Both Hugging Face repository cards include the current teaser, method overview, and HC-VAE figures under stable asset names.