qizhangslam/blur-aware-learn2splat-prism3d

Blur-Aware Learn2Splat: PRISM3D artifacts

Public engineering artifacts for the blur-aware Learn2Splat pipeline at code commit 74f55e6. The matching source branch is agent/blur-aware-cross-dataset.

Publication visualizations

Use the figures under publication_visuals/raw_blurred_input_vs_ours_10k/ for qualitative comparison. Every row is an exact camera/frame pair: the left column is the real RAW blurred input and the right column is our rendered deblurred output. No sharp GT, hold image, Turtle/EVSSM teacher, or alternate checkpoint is displayed as the input.

Complete visualization archive

archive/all_prism3d_visual_artifacts/ mirrors every PNG, JPEG, JSON, and CSV visual/evaluation artifact currently present under the local PRISM3D experiment roots. It contains 605 files, including 172 BPN convolution-kernel images and kernel-statistic tables, plus hold renders, RAW/output figures, longitudinal plots, receipts, and diagnostic comparisons. Original experiment-root names and relative paths are preserved.

This archive intentionally includes exploratory and rejected ablations. It is provenance evidence, not a claim that every archived result is accepted. Use only publication_visuals/ for the clean paper-facing RAW-to-Ours figures and accepted/ for admitted checkpoints and metrics.

Intermediate variables

intermediate_visuals/prism3d_10k/<scene>/ contains four measured views for each of the eight scenes:

• raw_turtle_teacher_ours.png: RAW input, the actual Turtle step-24K teacher, and our output at matching frame/camera indices.
• raw_evssm_baseline_ours.png: the same RAW and output with EVSSM shown only as a comparison baseline. EVSSM was not the teacher for these accepted runs.
• latent_3d_gaussian_scene.png: three PCA-aligned projections of the colored Gaussian centers from the saved PLY; this is not a camera render.
• gaussian_capacity_dynamics.png: measured primitive count, clone/split/prune actions, and blur-quality reward from the training receipt.

Accepted eight-scene 10K result

Scene	PSNR	SSIM	LPIPS	Primitives
bench	37.313	0.9259	0.0405	379,219
camellia	36.486	0.9347	0.0275	615,891
dragon	40.014	0.9647	0.0233	450,552
jars	38.625	0.9484	0.0304	488,866
jars2	36.641	0.9065	0.0775	360,193
postbox	36.587	0.9298	0.0420	516,129
stone_lantern	38.534	0.9489	0.0544	239,166
sunflowers	38.159	0.9373	0.0754	212,988
Average	37.795	0.9370	0.0464	407,876

accepted/10k/<scene> contains the point cloud, blur-objective state, receipt, diagnostics, BPN visualization, hold visualization, and a qualitative RAW-blurred-input/output comparison for each scene.

Accepted 50K continuations

accepted/50k contains longitudinal checkpoints for jars2 and sunflowers. The best hold PSNR is 37.1166 at 40K for jars2 and 40.6213 at 50K for sunflowers. The receipts preserve every scheduled 10K--50K metric.

The diagnostic fixed strong-blur sunflowers checkpoint comparison is in summary/sunflowers_raw_10k_base50k_lap02_50k_comparison.png. Its leftmost column is the RAW blurred input; it demonstrates that aggregate hold metrics can improve while difficult blurred views regress after 10K.

Negative ablation

ablations/rejected_lap02_50k increases only the Laplacian weight from 0.1 to 0.2. At 50K it reaches 40.3251 PSNR / 0.9536 SSIM / 0.0542 LPIPS, below the accepted 0.1 run on all three quality metrics. It is retained as negative evidence and is not the default pipeline.

Checkpoint scope and integrity

point_cloud.ply plus blur_aware_objective.pt are inference/render artifacts. They do not contain optimizer state and are not resumable training checkpoints. FILE_MANIFEST_SHA256.tsv is the authoritative path, size, and SHA-256 inventory for this release. Dataset images and the Turtle checkpoint are not redistributed here.