SplatAtlas-IR: Task-Aligned Evaluation of 3DGS Inverse Rendering for Robot Contact

Release status: Paper under review · Code to be released · Experiment artifacts available in this repository

The numbers, task definitions, and figure on this card match the submitted manuscript. Dated snapshot packages in the repository retain earlier experiment archives for provenance; they should not be read as a replacement for the protocol below.

TL;DR

SplatAtlas-IR evaluates 11 3D Gaussian splatting (3DGS) inverse-rendering (IR) methods on 46 scenes from 6 datasets, combining property measurements with a shared robot-contact test. Property evaluation covers novel-view synthesis (NVS), albedo, roughness, normals, and relighting; NVS, albedo, and relighting produce different method rankings. For geometry, a controlled simulation uses depth to locate the contact point and normals to orient the tool and set its approach direction. The protocol is developed on four objects, then frozen for evaluation on eight additional objects. Nine native-normal interfaces achieve 41.02–83.98% success at a 10 mm tolerance on the same 256 confirmation targets. Invalid inputs and execution failures are counted as failures. Six non-IR geometry interfaces provide external comparisons: vanilla 3DGS reaches 77.34%, while ObjSplat with RGB-derived multi-view stereo (MVS) depth reaches 95.31%. Ground-truth point and normal substitutions further show how the two geometry inputs affect contact outcomes.

Rendering quality alone does not establish whether a robot can use these outputs.

Project overview

3DGS IR methods recover different combinations of geometry, materials, and illumination. Published scores are difficult to compare because scene subsets, masks, output conventions, aggregation rules, and baseline implementations vary. SplatAtlas-IR re-runs the methods with shared data preparation and property-specific evaluation, then evaluates recovered depth and normals with one contact controller.

Contributions

1. A benchmark resource. Common evaluation protocols and a multi-terabyte collection of outputs, per-scene metrics, and records, publicly available in this repository.
2. An evaluation of recovered properties. Measurements of rendering quality, albedo and roughness accuracy, and normal accuracy and coverage, with comparisons across evaluation tracks.
3. A registered-pixel contact test. A fixed controller, point/normal substitutions, workcell sensitivity tests, and six non-IR interfaces connect recovered geometry to contact outcomes.

This is an artifact archive, not a source-code or raw-dataset release. Upstream method repositories, Conda environments, original datasets, caches, and sensitive files are intentionally excluded.

Important: the existence of an archive means that an attempt was preserved; it does not by itself mean that the run completed successfully or was included in the final comparison. Use the status tables below together with the manifests and recorded metrics.

Repository layout

Path	Contents
outputs/	Flat per-method/per-scene result archives. Suffixes such as _depth.tar, _events.tar, _model.tar, and _renders.tar identify the archived artifact family where that family exists. Not every method emits every family.
snapshots/2026-08-27/	Early result-freeze manifests and exclusion audit.
snapshots/2026-08-28/	Packaged result snapshot with package and exclusion manifests.
snapshots/2026-08-29-v3/	Method → scene formal benchmark archive, plus auxiliary results and scene/package manifests.
snapshots/2026-09-02-ycb-contact9/	Additional YCB experiment outputs, manifests, shared adapters, and supporting evidence.

Snapshot lineage

• 2026-08-27 is an early manifest/exclusion freeze.
• 2026-08-28 adds the corresponding packaged result snapshot.
• 2026-08-29-v3 reorganizes the formal benchmark as method → scene → semantic artifact packages and stores auxiliary results separately.
• 2026-09-02-ycb-contact9 adds later YCB normal, contact, adapter, and supporting experiment outputs.
• outputs/ provides the flat method/scene access layer and may contain later recovery runs not represented by an older dated summary.

Dated snapshot archives are uncompressed POSIX tar files. Their paths are preserved relative to the original experiment root under /root/autodl-tmp. Historical packages may include analyses that are not reported in the manuscript (for example, earlier roughness-ordering or pushing diagnostics). Treat those packages as provenance, not as the paper protocol.

For the v3 formal archive, a scene should be treated as archive-complete only when its scene_manifest.json records archive_complete=true and the stored package sizes and SHA-256 hashes match. The YCB snapshot uses the analogous unit_manifest.json gate.

Useful audit entry points include:

• snapshots/2026-08-29-v3/manifests/manifest_summary.json
• snapshots/2026-08-29-v3/manifests/scene_manifest.jsonl
• snapshots/2026-08-29-v3/manifests/package_manifest.jsonl
• snapshots/2026-09-02-ycb-contact9/manifests/unit_manifest.jsonl

Artifact schema

The v3 and YCB snapshots group one logical run into up to four semantic packages:

Package family	Typical contents
01_analysis_evidence.tar	Aggregated metrics, audit evidence, validity records, and analysis outputs.
02_metadata_logs.tar	Method configuration, command metadata, runtime logs, and scalar trajectories.
03_model_geometry_state.tar	Gaussian fields, checkpoints, point clouds, or other method-native geometry/state.
04_renders_predictions.tar	Rendered RGB, decomposition channels, relighting outputs, normals, and task-facing predictions where available.

Some units are split into numbered parts to stay below the package-size target. The exact package list, source-relative path, source byte count, file count, remote byte count, SHA-256, and completion state are recorded in the corresponding manifest. A method that does not natively produce a channel will not contain a synthesized substitute for that channel.

Paper-to-artifact map

Paper component	Primary artifact family	Validity/provenance record
NVS and albedo	Render/prediction packages and per-scene analysis evidence	Scene manifest, method metadata, metric records
Roughness	Native roughness maps compared on matched pixels	Channel identity, export decoding, pixelwise MAE records
Normal benchmark	Native normals or separately labeled depth-derived proxies	Source type, angular-error records
Common relighting	Environment-map render outputs for the five executable interfaces	Environment-map identity, scene/view intersection, metric records
Registered-pixel contact	Frozen target registry, source depth/normals, planning/execution records, contact outcomes	D4 development / C8 confirmation split and source–target result records

The exact file path for a reported value should be recovered through the manifest rather than guessed from a tar filename.

Benchmark scope and source datasets

The common property benchmark contains 46 scenes across six scene groups:

Scene group	Scenes
NeRF Synthetic	8
Shiny Blender	6
Stanford-ORB inverse-rendering scenes	9
Mip-NeRF 360	9
Tanks and Temples	12
Deep Blending	2
Common total	46

The scene selection is fixed as follows:

• the complete NeRF Synthetic, Shiny Blender, and Mip-NeRF 360 sets;
• all three Stanford-ORB lighting conditions for each of chips, cup, and pitcher;
• the 12-scene Tanks and Temples subset used by NerfBaselines: Barn, Caterpillar, Truck, Lighthouse, Playground, Train, Auditorium, Ballroom, Courtroom, Museum, Palace, and Temple; and
• the two Deep Blending scenes used in the original 3DGS evaluation: Dr Johnson and Playroom.

GaussianShader also has an archived bathroom diagnostic run. It is outside the 46-scene common aggregate.

Raw source datasets are not redistributed here. Users should obtain them from their official providers and comply with the corresponding licenses and terms.

Execution protocol

We run the property benchmark using the default settings in the public implementations on an NVIDIA RTX 4080 SUPER. Data preparation, splits, output checks, and evaluation metrics are standardized while retaining the recommended training recipes. Unavailable channels are not fabricated. Native normals are kept separate from depth-derived proxies. Failed results are not imputed.

For contact, each object is trained separately with 100 training views and evaluated on 25 test views. Depth is converted to linear distance along the normalized camera ray in the common render frame. The controller reads the exact target pixel, without interpolation, inpainting, or nearest-valid substitution.

Evaluation tracks

Test	Outputs evaluated	Test set	Measures
NVS / albedo	Rendered RGB and albedo	46 scenes; property-specific subsets	PSNR, SSIM, LPIPS; SI-PSNR
Roughness	Native roughness maps	7 outputs; 4 objects, 200 views/object; matched GGX regions on Ficus and Hotdog	Pixelwise MAE; field comparison
Normals	Native normals and depth-derived proxies	NeRF Synthetic: 800 views/source. YCB: 12 contact objects, 25 views/object	Angular MAE
Relighting	Renders after environment-map swap	5 interfaces. NeRF Synthetic: four GT scenes, 8 lights. Stanford-ORB: 180 images/method	PSNR
Contact	Depth and normal interfaces: 9 IR, 6 non-IR	D4: 4 objects, 128 targets. C8: 8 objects, 256 targets	SR@10; 95% CI; point/normal controls

Method output capabilities

Method	Year	Explicit roughness	Normal source	Metalness	Paper reports relighting	Validated common env-map run
GS-IR	2024	✓	Native	✓	✓	✓
RTR-GS	2025	✓	Native	✓	✓	✓
SVG-IR	2025	✓	Native	—	✓	✓
Ref-Gaussian	2024	✓	Native	—	✓	✓
Normal-GS	2024	—	Native	—	—	—
GaussianShader	2024	✓	Native	✓	✓	—
GI-GS	2025	✓	Native	—	✓	✓
3DGS-DR	2024	—	Native	—	—	—
Spec-Gaussian	2024	—	Depth-derived proxy	—	—	—
3iGS	2024	—	Depth-derived proxy	—	—	—
DiscretizedSDF	2025	✓	Native	—	✓	—

“Paper reports relighting” records the method paper's capability claim. “Validated common env-map run” is stricter: GaussianShader and DiscretizedSDF report relighting in their papers, but the released code does not expose an environment-map-swap interface for this protocol. Five methods are therefore evaluated for relighting.

Results at a glance

Novel-view synthesis

Scene-level scores are weighted equally. SD is the sample standard deviation of scene-level PSNR.

Method	PSNR ↑	SD	SSIM ↑	LPIPS ↓
GS-IR	21.50	4.29	0.809	0.246
RTR-GS	25.94	7.18	0.849	0.216
SVG-IR	21.50	4.18	0.744	0.343
Ref-Gaussian	27.20	6.53	0.852	0.218
Normal-GS	27.26	6.27	0.875	0.190
GaussianShader	24.71	8.03	0.814	0.246
GI-GS	21.62	11.21	0.710	0.241
3DGS-DR	27.65	6.73	0.861	0.198
Spec-Gaussian	22.43	7.15	0.838	0.225
3iGS	28.14	6.43	0.877	0.170
DiscretizedSDF	26.64	10.40	0.852	0.172

Matched NVS and relighting

For the five methods in the common relighting test, evaluated scenes are matched within each dataset and NeRF Synthetic and Stanford-ORB are given equal weight. Ref-Gaussian has the highest NVS score and the lowest relighting score in this comparison.

Method	NVS ↑	Relighting ↑
GS-IR	19.57	25.50
GI-GS	15.60	22.72
SVG-IR	22.34	22.58
RTR-GS	27.71	26.54
Ref-Gaussian	30.29	22.30

Albedo and relighting

Albedo uses scale-invariant PSNR (SI-PSNR). On NeRF Synthetic, one least-squares RGB scale per method–scene pair is shared across all 200 views. ORB-A uses Stanford-ORB pseudo-GT; the NeRF Synthetic columns use Drums, Ficus, Hotdog, and Lego; ORB-R follows the official background-inclusive protocol (180 images/method). GaussianShader has an albedo output but is not part of this quantitative comparison.

On the NeRF Synthetic scenes shared by the five relighting methods, RTR-GS ranks fourth in albedo and first in relighting, while GI-GS ranks second and fourth, respectively.

Method	Albedo ORB-A	Albedo NeRF Synthetic	Relighting NeRF Synthetic	Relighting ORB-R
GS-IR	17.54	17.32	20.17	30.83
GI-GS	16.50	17.24	15.92	29.51
SVG-IR	16.53	14.94	18.82	26.34
RTR-GS	12.51	13.73	22.07	31.00
Ref-Gaussian	12.56	12.81	13.50	31.09
DiscretizedSDF	17.21	13.66	—	—

Roughness

Roughness is compared at the same pixels of all 200 test views of Drums, Ficus, Hotdog, and Lego, using reference fields rendered from the official NeRFactor Blender assets. Quantitative evaluation uses regions with one distinct active isotropic GGX or Multi-GGX reflection-roughness signal (α = r²), without transmission or glass coupling. These regions occur in Ficus and Hotdog. Drums and Lego contain Beckmann and multi-branch materials; their fields are retained for spatial comparison rather than combined into a scalar reference.

Native export decoding uses 0.96 r_raw + 0.04 for GS-IR, the native range for GI-GS, and opacity unpremultiplication followed by the export clamp for DiscretizedSDF. RTR-GS uses direct roughness. No fitted calibration is applied.

MAE is computed pixelwise within each view and then averaged equally across the 200 views. Eligible regions cover 22.24% of Ficus and 91.84% of Hotdog foreground pixel instances across all views; all seven outputs are valid on this common support. Ref-Gaussian has the lowest MAE on both objects; errors span 0.115–0.495 on Ficus and 0.214–0.514 on Hotdog.

Method	Ficus ↓	Hotdog ↓
GS-IR	0.329	0.504
RTR-GS	0.416	0.274
Ref-Gaussian	0.115	0.214
SVG-IR	0.186	0.372
GI-GS	0.495	0.435
GaussianShader	0.395	0.413
DiscretizedSDF	0.305	0.514

Normals

Normal accuracy is the mean angular error between predicted and reference unit normals.

On 800 NeRF Synthetic views per source, native-output MAE ranges from 23.74° to 43.13°; the range extends to 57.22° when depth-derived proxies are included.

Method	Source	MAE (°)
GS-IR	native	34.71
RTR-GS	native	24.25
SVG-IR	native	43.13
GI-GS	native	35.55
GaussianShader	native	33.44
Ref-Gaussian	native	23.74
DiscretizedSDF	native	30.83
Normal-GS	native	24.85
Spec-Gaussian	proxy	57.22
3DGS-DR	native	25.99
3iGS	proxy	52.15

On the same 12 YCB contact objects (25 views/object), failure-aware (FA) MAE assigns invalid predictions 180°; valid-only (VO) MAE averages valid predictions only. Coverage is the object-mean valid-prediction percentage. MAEs are averaged equally over views and then over objects.

Method	FA MAE	VO MAE	Cov. (%)
DiscretizedSDF	7.04	7.04	100.00
GI-GS	16.09	16.09	100.00
GS-IR	17.19	17.19	100.00
3DGS-DR	15.50	12.70	97.88
RTR-GS	7.27	7.22	99.97
Normal-GS	21.75	21.60	99.90
Ref-Gaussian	27.77	5.07	86.80
SVG-IR	30.61	30.61	100.00
GaussianShader	60.43	42.82	88.97

Registered-pixel contact

A 7-DoF Panda touches specified image targets in a known, registered MuJoCo scene. Depth gives the 3D point along the camera ray; the supplied normal sets the inward tool axis and approach from an 80 mm standoff.

D4 (development) contains gelatin box, pudding box, pear, and orange. C8 (confirmation) contains golf ball, apple, baseball, mustard bottle, bleach cleanser, tomato soup can, cracker box, and bowl. Each contributes 32 targets. Controller settings, the base library, and its selection rule are chosen on D4 and frozen before reading C8 outcomes. Neither method outputs nor contact outcomes enter target selection; all 384 targets are retained.

SR@10 requires every object-side surface contact point at the first active pad–object contact to lie within 10 mm of the registered target. Invalid inputs, planning failure, no contact, wrong-part contact, and off-target contact are failures. Primary success rates use 95% confidence intervals from 5,000 bootstrap draws, sampling objects and then targets within each sampled object.

The primary condition, (P_m N_m), uses the predicted point and normal. (P_0 N_m) replaces the point with the registered GT collision point; (P_m N_0) replaces the normal with the registered GT normal. All conditions use the same 256 C8 targets. The shared (P_0 N_0) control succeeds on all 256.

With both point and normal predicted, success ranges from 41.02% to 83.98% across nine IR methods. High success with one GT input need not carry over when both inputs are predicted: Ref-Gaussian reaches 98.05% with a GT normal and 72.66% with a GT point; DiscretizedSDF exceeds 95% in both single-oracle conditions but is 83.98% when both inputs are predicted.

Method	(P_0 N_m) (GT point)	(P_m N_0) (GT normal)	(P_m N_m) (both predicted)	95% CI
DiscretizedSDF	96.88	95.31	83.98	[69.14, 94.53]
GI-GS	91.80	96.48	81.64	[69.14, 91.02]
GS-IR	92.19	95.31	76.17	[63.28, 86.72]
3DGS-DR	87.89	90.63	73.83	[61.72, 84.77]
RTR-GS	95.70	93.36	73.05	[58.20, 85.94]
Normal-GS	88.28	87.89	66.80	[52.73, 78.91]
Ref-Gaussian	72.66	98.05	65.63	[48.83, 82.81]
SVG-IR	71.09	70.31	47.66	[31.25, 63.29]
GaussianShader	48.05	69.14	41.02	[25.39, 57.03]

Invalid inputs and execution failures remain in the denominator.

Six non-IR geometry interfaces use the same C8 targets, controller, and (P_m N_m) endpoint. MVS-D uses MVSFormer++ depth estimated from the input RGB; Mesh-D uses depth rendered from the known object mesh. Vanilla 3DGS reaches 77.34%, within the IR range. ObjSplat MVS-D reaches 95.31%, exceeding DiscretizedSDF by 11.33 percentage points (paired 95% CI [3.13, 22.66]).

Geometry interface	Training supervision	SR@10	95% CI
3DGS depth-derived	RGB	77.34	[54.69, 93.36]
ObjSplat MVS-D	RGB + RGB-derived MVS depth	95.31	[89.84, 99.22]
DreMa MVS-D	RGB + RGB-derived MVS depth	82.42	[74.22, 89.84]
GaussianGrasper	RGB + mesh-derived depth	69.53	[60.16, 78.12]
ObjSplat Mesh-D	RGB + mesh-derived depth	96.48	[93.36, 98.83]
DreMa Mesh-D	RGB + mesh-derived depth	90.23	[84.38, 95.70]

Two workcell variants, chosen without reference to outcomes, fix the radius at 0.55 m and retain eight azimuths, with heights fixed at 0.42 m and 0.58 m. Across 18 comparisons (two variants and nine IR methods), the largest absolute change in SR is 1.95 percentage points, and all paired 95% CIs include zero.

A separate retrospective (P_0 N_m) analysis on 119 retained targets (69 in C8) finds a stronger association between target-local normal error and success than whole-view error. Entries below are (-\rho), where (\rho) is Spearman's rank correlation.

Set	Whole-view	Target-local	Δ [95% CI]
All12	0.603	0.781	0.178 [0.027, 0.387]
C8	0.438	0.730	0.291 [0.066, 0.572]

Methods retained in the final property benchmark

“Run coverage” records the artifact inventory used to organize this repository. It is not inferred merely from archive presence, and a covered scene may still have a modality-specific validity restriction.

Method	Archive key	Run coverage	Reporting note
Normal-GS	normal_gs	46/46	Full common-scene coverage.
GaussianShader	gaussianshader	46/46	Full common-scene coverage; bathroom is an extra diagnostic scene.
GI-GS	gigs	46/46	Full common-scene coverage after enabling a valid PBR training phase.
SVG-IR	svg_ir	46/46	Full common-scene coverage; only validated outputs are used in aggregate reporting.
3DGS-DR	three_dgs_dr	46/46	Full common-scene coverage after upstream compatibility fixes.
3iGS	three_igs	46/46	Full common-scene coverage after recovery from early outdoor-scene OOM runs.
Spec-Gaussian	spec_gaussian	46/46	Full common-scene coverage; the tested implementation does not expose a roughness output.
DiscretizedSDF	dsdf	26/46	Partial accepted coverage; invalid zero-Gaussian runs are excluded.
Ref-Gaussian	ref_gaussian	46/46	Full common-scene coverage after an indirect-stage null guard.
RTR-GS	rtr_gs	46/46	Full common-scene coverage; difficult scenes can have high densification and cost.
GS-IR	gs_ir	46/46 outputs	PBR novel-view results are accepted only for four Blender scenes; remaining standard RGB outputs are from the SH/geometry stage and are not reported as PBR NVS.

DiscretizedSDF's 26 accepted scenes consist of the object-centric coverage plus three accepted real-scene runs. The excluded portion contains 8 Mip-NeRF 360 scenes, 10 Tanks and Temples scenes, and both Deep Blending scenes.

Methods not retained and why

These exclusions are based on the tested upstream implementations at the evaluated scene scale. They are not claims that the underlying research ideas are invalid, and later upstream revisions may behave differently.

Method	Accepted evidence	Method-side issue observed	Exclusion decision
GIR	5 Tanks and Temples scenes	Densification grew to roughly four million Gaussians. The resulting nvdiffrast workload exceeded a CUDA grid dimension during rasterization; a UV-grid compatibility patch removed one launch-shape error but did not solve the underlying scale growth.	Excluded because coverage remained too small and the implementation was not stable at the benchmark's scene scale.
GS-ROR	No valid final scene	The tested implementation assumes an object-centric SDF AABB of approximately [-1.5, 1.5]^3. On large real scenes (one observed extent was approximately 82 × 13 × 78), optimization became invalid or degenerate, including negative loss, infinite PSNR, and poor test renders. SDF initialization was also prohibitively slow.	Archived attempts are preserved, but none were accepted into the final comparison.
Relightable3DGaussian (R3DG)	9 scenes	The upstream path requires masks and initially contained an image/mask broadcasting mismatch. After input-contract repair, full-resolution images and visibility structures still consumed about 78 GB of host RAM on larger scenes and could stall the host.	Excluded because only 9 scenes were accepted and the remaining scale failure was unresolved.
IRGS	1 Blender scene	The implementation traces 256 secondary rays for every visible pixel. On the Chair scene this produced roughly 76 million rays and exceeded a 32 GB GPU. Subsampling through train_ray would materially change the original training behavior.	Excluded because the faithful configuration was not scalable and only one scene was accepted.

Considered but not evaluated

Mani-GS reached only preliminary integration. No validated benchmark run was produced, so it is not labeled as a runtime failure and is not included in the ranking.

Method-side failures and limitations observed in retained methods

The table below records reproducible behavior in the tested method revisions. Benchmark-orchestration incidents are intentionally out of scope.

Method	Observed behavior	Treatment in this release
Normal-GS	One teapot run produced constant-white RGB output; the cause remained unresolved.	The affected output is flagged and is not used as evidence of successful rendering quality.
GaussianShader	On high-detail Stanford-ORB training, a 30k run densified to about 3.19 million Gaussians and developed needle-like geometry.	The method remains included, but scale-sensitive runs are interpreted with this failure mode in mind.
GI-GS	The tested default schedule placed pbr_iteration at the total training iteration, leaving no effective PBR phase. After a corrected PBR run, albedo remained exactly achromatic (R=G=B) and metallic values stayed at an activation boundary in the inspected outputs.	Results come from a run with an actual PBR phase; the observed material degeneracy is retained as a limitation rather than hidden.
SVG-IR	Several runs showed high memory/runtime cost, collapsed object albedo, roughness saturation near 1, poor normals, or unstable crystal-ball-style relighting.	Only completed, validated artifacts are used; these output degeneracies are disclosed.
3iGS	Early outdoor-scene runs exhausted GPU memory.	Later recovered runs completed the common coverage; the early OOM attempts remain archived as provenance.
Spec-Gaussian	Early outdoor-scene runs exhausted GPU memory or became substantially slower. The tested code does not provide a roughness channel.	Recovered runs are used where valid; no roughness result is attributed to the method.
DiscretizedSDF	On many real/Colmap scenes, pruning removed all Gaussians. Symptoms included zero loss, infinite PSNR, and downstream reshape failures with N=0.	Those runs are invalid, not successes. Only 26 accepted scenes are reported in the archive inventory.
Ref-Gaussian	Inspected outputs showed roughness saturation near 0, an environment map clamped near 1, and visible banding in some relighting results.	Valid runs remain included, with these material/lighting limitations disclosed.
RTR-GS	The method has a high fixed computational cost and can densify heavily on difficult Stanford-ORB scenes.	Completed runs remain included; compute and scale behavior should be considered when comparing practicality.
GS-IR	Inspected material outputs saturated at roughness 1 and metallic 0. Only four Blender scenes yielded valid PBR NVS results; other standard RGB renders were produced by the SH/geometry stage.	PBR and non-PBR outputs are reported separately and are not conflated.

Minimal upstream compatibility changes

The following local changes were made to allow the tested upstream code paths to execute. They were limited to loader contracts, shape handling, debug statements, null safety, or intended phase activation. Benchmark scoring rules were not changed by these patches.

Method	Local change	Scope
Normal-GS	Supplied neutral placeholder normal files because the Synthetic loader unconditionally required them even though the loaded field was not consumed by the tested run path.	Loader compatibility only.
3DGS-DR	Initialized HWK on loader paths where it was otherwise unbound; derived intrinsics from the actual resized image dimensions to remove an off-by-one reshape mismatch; added .JPG/.jpg filename compatibility.	Camera/image loading compatibility; no objective change.
3iGS	Removed a leftover interactive pdb/st() breakpoint from the upstream execution path.	Non-interactive execution only.
GI-GS	Moved the PBR-phase start before the final iteration so that the intended PBR stage actually ran.	Training-schedule correction; no evaluator change.
DiscretizedSDF	Guarded or removed a debug mask.save("test.png") call that attempted to save a floating-point PIL mode-F image as PNG.	Debug I/O only.
Ref-Gaussian	Added a guard for extra_dict is None in the indirect-light stage.	Null safety only.
GIR (excluded)	Reshaped a one-dimensional UV query into a two-dimensional launch grid to avoid a CUDA grid-dimension overflow.	Compatibility attempt; it did not resolve densification-driven scaling.
R3DG (excluded)	Expanded 2-D masks to H × W × 1 and supplied no-occlusion masks where the upstream loader required them.	Input-contract compatibility; it did not resolve host-memory scaling.

Unless stated above, failed attempts were not converted into successful results through evaluator changes, silent metric substitution, or reduced-fidelity method variants.

Downloading selected artifacts

This repository contains large tar archives. Download only the method/scene files you need instead of cloning the entire repository.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="KCBtheone/splatatlas-ir",
    repo_type="dataset",
    allow_patterns=["outputs/normal_gs_Chair_*.tar"],
    local_dir="splatatlas-ir",
)

Extract an archive into a separate inspection directory:

mkdir -p extracted
tar -xf splatatlas-ir/outputs/normal_gs_Chair_depth.tar -C extracted

Because this is an artifact repository rather than a row-oriented dataset, use huggingface_hub or the Hugging Face file browser rather than expecting datasets.load_dataset() to construct a conventional train/test table.

Intended uses

This release is intended for:

• inspecting the experiment evidence behind SplatAtlas-IR;
• reproducing or auditing property-specific NVS, albedo, roughness, normal, and relighting comparisons;
• verifying method/scene coverage, failed attempts, output provenance, and package integrity;
• studying how native versus depth-derived properties behave under fixed evaluation interfaces; and
• reproducing the reported registered-pixel contact diagnostic once the accompanying evaluation code is released.

Out-of-scope uses

This repository should not be used as:

• a replacement for the original training datasets or upstream method repositories;
• a conventional supervised-learning dataset or a source of ground-truth physical material parameters;
• evidence that one method is universally best across datasets, properties, or downstream tasks;
• evidence of end-to-end robot perception, global motion planning, grasping, or real-robot manipulation; or
• evidence that a failure observed in the tested revision persists in later upstream revisions.

Limitations

• Property availability is uneven. Methods expose different channels, and only five provide validated common environment-map relighting runs. Missing outputs are not synthesized.
• Coverage is uneven. DiscretizedSDF has 26 accepted scenes in the archive inventory, and failed method–scene cells are not imputed.
• Protocol-specific references differ. Stanford-ORB albedo uses pseudo-GT, while NeRF Synthetic uses synthetic GT; results are compared within a dataset, not across these reference types.
• Hardware and implementation revisions matter. The reported runs use tested public implementations on an RTX 4080 SUPER. Runtime, memory behavior, and compatibility failures may differ on other revisions or hardware.
• The contact evaluation is simulated and registered. It uses known target correspondence, a fixed Panda/controller/tool, and one deterministic execution per method–target pair. D4 is used to freeze the protocol; C8 is the confirmation set. It is not an end-to-end robotics benchmark.
• Some archived units are intentionally unsuccessful. Manifests, validity flags, and the manuscript tables—not archive presence—determine whether a result is usable.

The release contains generated experiment artifacts from synthetic or publicly available benchmark scenes and simulated robotics experiments. It is not intended to contain personal or sensitive information, and the original datasets are not redistributed.

Reproducibility and interpretation

• Consult the exact archived configuration, logs, and method metadata before reproducing a run; method defaults and upstream code may change.
• Verify package byte size and SHA-256 against the relevant manifest before treating a local download as complete.
• Treat archive_complete, metric validity, and final inclusion as separate checks.
• Failed and partial runs are retained when useful for provenance. Their presence must not be interpreted as a positive result.
• Comparisons should respect each method's native output contract. In particular, a missing material channel is not synthesized and a geometry-stage RGB render is not relabeled as PBR output.

License and third-party material

The benchmark-authored documentation, manifests, and released result packaging are provided under CC BY 4.0. Upstream method names, papers, software, and source datasets remain governed by their respective owners and licenses. This repository does not redistribute the upstream source repositories or original benchmark datasets; obtain those materials from their official sources and follow their terms.

Citation

Citation metadata will be added when the accompanying SplatAtlas-IR manuscript is publicly released.

Contact

During anonymous review, questions may be raised through the Hugging Face repository's Community tab. Author names, institutional details, and a permanent contact address will be added after the review period.