Lee1413/SimAnything-SAGE-train-debug

SimAnything SAGE whole-room debug cases

This release corrects the previous chunk-only delivery. The six canonical entries in manifest.jsonl are now whole-room cases (<room>_fullscene). Both GT and predictions have no final chunk clipping, and roi is null.

These are training rooms seen by the upstream model. They are train_debug, excluded from unseen-room/generalization and standard repair benchmark claims. No editing-agent runs or paid calls are included.

GT is the complete original evaluated Blender room (all exported faces and instance labels), including room structure and furniture. The original arrays are also provided losslessly as gt/original_room.npz. The prediction is freshly reconstructed across the full RGB-inferred scene domain with automatic multi-chunk A4 + N3 inference. Full-domain inference does not guarantee complete or accurate surfaces; the table reports whole-room recall explicitly.

Whole-room baseline

Metrics use the packaged editing input (at most 300,000 faces), against the complete GT: 200,000 area-weighted samples per mesh, seed 42, meshedit.eval.v1, no ROI. Chamfer is in metres; normal consistency uses the existing 20 cm correspondence cutoff.

Room	Inferred chunks	Full GT faces	F@5cm	Precision@5cm	Recall@5cm	Chamfer (m)	Normal consistency
11ca9d33	9	367,445	0.3294	0.2596	0.4508	0.1993	0.5593
4ab80e99	6	678,420	0.4690	0.5125	0.4323	0.1620	0.6845
82e428e2	9	1,288,053	0.5485	0.4957	0.6139	0.0775	0.7306
8a0655e3	6	497,640	0.3693	0.3258	0.4260	0.1700	0.5676
a6b5d2b6	16	13,990	0.4099	0.3520	0.4905	0.1700	0.7741
ebd1ece9	9	1,496,373	0.6683	0.6279	0.7143	0.0494	0.8345

The old high SS IoU / F-scores were selected-chunk scores, not whole-room scores. They cannot be compared directly with this table. Whole-room SS IoU has not been established on a common occupancy grid and is not claimed.

Visualizations

Top views show prediction and full GT at the same scale. Points at or above the GT ceiling height minus 5 cm are hidden only in these overview images to expose the interior; downloadable meshes and evaluation retain every surface. GT and prediction overlays for all 16 reference views are under previews/ (producer QA, never agent inputs).

Inputs, scale and depth

• Six original rooms, each with 16 RGB frames and cameras in original room coordinates.
• Inference uses RGB-derived VGGT cameras/depth and automatic chunking; no GT meshes, chunk transforms or GT depth enter reconstruction.
• After inference, a single camera-centre Sim(3) maps the entire output into SAGE room coordinates. No mesh-to-GT registration. GT camera centres supply metric scale; this is not a scale-recovery benchmark.
• The agent input is globally decimated to at most 300k faces, without spatial cropping. The unmodified full-resolution reconstruction is retained in raw_predictions/, with separately computed raw whole-room surface metrics in provenance. The supplementary raw report omits expensive topology diagnostics; its geometric scoring is numerically cross-checked against the unchanged canonical baseline.
• Five cases include the original 32×32 VGGT patch depth reprojected into the case cameras. 82e428e2_fullscene uses RGB/cameras only because v024 fails whole-room depth agreement (all depth omitted, no GT-based pixel filtering). Every room has a retained depth preflight report.
• 11ca9d33 now passes depth agreement against complete GT; the previous chunk GT omitted occluding surfaces and its local rejection does not apply to the whole-room check.
• Only inputs/, the instruction and units go to the editing agent. GT, baseline, raw predictions, QA and provenance are withheld.

Frozen method

VGGT-Omega + A4 SS RAW step11980 + N3-C Shape RAW step25 + frozen FDG decoder, automatic full-scene MultiDiffusion. Seed 42, 16 normalized RGB inputs, 512 pipeline, raw-index-crop-v1, sage-validation SS sampler, no validation crop and no final clipping. The complete commands and runtime configuration are recorded per room.

• SimAnything: 6bd60214c13c670b83838fd467eae3179944aea2.
• MeshEdit: 207e318908d7553c9216547d2b2571cd3e3b9d92.
• SS SHA-256: 03d1e1b1e2a3b86b9a312d193be549c926191dc66e7c2f10bcee5419bb349e31.
• Shape SHA-256: db07a7c57c84247b80fb93d34f1b3820ea1726316c40911f7ed4c67da63b6d96.
• VGGT SHA-256: c02da418b18bb01d0392598d3f6147366bcde1bb70fd08a5e3bf7925b0667934.

Download

hf download Lee1413/SimAnything-SAGE-train-debug --repo-type dataset --local-dir sage_train_debug
meshedit case validate sage_train_debug/ebd1ece9_fullscene --verbose

Use the root manifest.jsonl to enumerate the six canonical full-room cases. A pre-existing local download may retain obsolete files; follow the manifest or download into a fresh directory.

Superseded chunk release

The entire previous payload from commit 22a491d2c1b037614c8cbd4e980535c79899bd59 is preserved byte-for-byte under legacy/chunks/, including its old manifest, card, GT, predictions and local metrics. It is historical chunk-only data, excluded from the root manifest. No original case data was discarded.

The private dev/test dataset Lee1413/SimAnything-SAGE-cases is unchanged. SAGE and embedded asset rights remain with their respective licensors; this private debug release grants no additional redistribution rights.