Phips/lucid-val-512

LUCID val 512 — 136 whole-scene validation tiles

Goal. A fixed, visually diverse, technically clean validation set for single-image super-resolution (SISR): 136 tiles, each from a different photo, each showing the whole scene rather than a random crop. Use it to validate during training and to compare SISR models against each other.

gt/   512×512  ground truth   (whole scene, see "How it was built")
x2/   256×256  bicubic LR     (Pillow BICUBIC)
x4/   128×128  bicubic LR     (Pillow BICUBIC)
val_scores.csv   — every tile's filter scores
contact_val.png  — 60 random tiles, for a fast eye check

Shard N in gt/, x2/, x4/ contains the same tile names, so zip(gt, x2, x4) yields matched HR/LR samples.

---

1. Source data

• UHD-IQA validation + test splits: 1,804 photographs at ~3840×2560 (4K), CC0 licensed.
• UHD-IQA is an image-quality dataset curated for human preference; its validation/test splits are disjoint from its training split.

2. How tiles were made

Per photo, exactly one tile:

1. Resize so the short side = 512 with Pillow BICUBIC (a = −0.5; the antialiased resampler, matching the MATLAB imresize convention used by standard SISR benchmarks). A 3840×2560 photo becomes 768×512.
2. Center-crop 512×512 — the frame keeps the full scene: a person with face, hair, clothing and background; a landscape with sky and ground.

Why not a random 1:1 crop from the 4K original? Two reasons, both measured: semantically, a 1:1 crop often captures "just hair" or "just grass", which is useless for judging whole-image SR quality; metric-wise, native 4K crops are so far outside the sharpness profile of quality models trained on ~512 px annotated images that they measure as degraded (topiq 0.31 vs 0.66 for the same photo as a whole-scene tile — see §4). The GT is therefore a single bicubic resample of the original — the standard whole-image benchmark convention — and the LR is a single further bicubic resample of that GT.

---

3. The filter pipeline (exact)

Every candidate tile passed six handcrafted signal gates (cheap, CPU, exact formulas below), one complexity gate, and four learned perceptual gates (neural, GPU). Gate thresholds were calibrated on a curated ground-truth reference (125 tiles of the PSISRD validation set) so that ~99 % of curated GT passes while measured degradation classes fail.

Signal gates (CPU, computed on the grayscale tile)

#	Gate	Threshold	Measures	Why
1	Entropy	≥ 5.2 bits	Shannon entropy of the 256-bin gray histogram	rejects flat/empty content (sky, walls) with no learning signal
2	Laplacian variance	80 < v < 10,000	variance of the CV_64F Laplacian	low = out-of-focus/defocus blur (wrong GT); high = extreme sharpening
3	Gradient energy	≥ 0.70	mean of (Sobel_x² + Sobel_y²), Sobel divided by 8	rejects textureless tiles
4	Blockiness	≤ 1.7	mean absolute gradient across JPEG 8×8 block boundaries ÷ (mean interior gradient + 1), computed as (boundary+1)/(interior+1)	JPEG compression blocks; ~1.0 for clean content, ~2.0 for Q30 JPEG flats
5	Aliasing ratio	≤ 0.80	mean FFT magnitude in the outer ring (radius ≥ 0.75·min/2) ÷ overall mean, after a Hann window	broadband high-frequency pollution: UI screenshots, text/stripes, ringing
6	Clipped ratio	≤ 0.40	fraction of pixels with value ≤ 2 or ≥ 253	blown/crushed frames carry unrecoverable signal

Complexity gate (GPU)

Gate	Threshold	Measures	Why
ICNet complexity (IC9600-style model, score calibrated per tile size; at 512 the calibration is identity)	≥ 0.30	learned density of structured detail	junk floor: removes sky/flat/defocus while retaining whole-scene composition

Learned perceptual gates (GPU, cascade — a tile failing one gate never reaches the next)

Gate	Threshold	Measures	Why
topiq_nr	≥ 0.665	no-reference quality, CFANet-Res50 trained on KonIQ-10k (real-world distortions)	catches JPEG damage, blur and visible noise that the handcrafted proxies miss
clipiqa+	≥ 0.45	CLIP-embedding quality probe	semantic plausibility; rejects odd/synthetic-looking content
niqe	≤ 6.5	Natural Image Quality Evaluator (lower = better), scene-statistics based	catches photographic grain and unnatural statistics; a visibly noisy tile measured 11.9 vs a 3.7–5.0 reference band
qrealign-mini	≥ 0.78	Q-Align-style VLM judge (0.8 B), quality task	final holistic verdict; the only metric that reliably sees oversharpening halos

Threshold provenance (the curated reference's own distribution): topiq_nr p01 = 0.665; niqe p95 = 6.65 (rejects 5.6 % of the reference); qrealign-mini p01 = 0.77; clipiqa+ = 0.45 keeps 100 % of the reference.

Deduplication

ResNet-18 ImageNet embeddings (512-D, L2-normalized); a tile is dropped when cosine similarity ≥ 0.96 with any tile already accepted. Comparison streams per tile (duplicates inside one processing batch cannot escape each other).

---

4. What happened to the 1,804 photos (measured funnel)

stage	photos	tiles
source photos	1,804	—
whole-scene 512 tile produced	1,804	1,804
tile passes all six signal gates	1,596	1,579
tile passes all perceptual gates + complexity	658	658
dedupe (0.96)	—	658 (no removals — one tile per photo)
diversity selection	136	136

Only 41 % of photos pass the strict gate stack: UHD-IQA is aesthetic-curated, not artifact-free, and the gates reject compression damage, grain, and over-processing. 12.5 % of the survivors are green-dominant (>30 % green pixels) — capped, not eliminated, because nature content belongs in a benchmark.

5. Selection for diversity (how the final 136 were chosen)

The candidate pool of 658 photos could pass every gate yet still be visually monotonous (e.g. many leaf close-ups). The final set maximizes visual spread:

1. One tile per photo — sibling crops of the same photo are near-duplicates.
2. DINOv2 ViT-B/14 embeddings (timm, semantic visual features).
3. Farthest-point (k-center) selection: start from a seeded random pick, then repeatedly add the tile that is farthest (lowest max cosine similarity) from everything already chosen.
4. Distinctness ceiling: reject any candidate with cosine similarity

   0.90 to an already-selected tile.

5. Mode caps: ≤ 5 tiles per k-means cluster (60 clusters over DINOv2 embeddings); ≤ 40 tiles per CLIP content category (15 broad categories: architecture, animal, vehicle, food, interior, portrait, …); ≤ 40 green-dominant tiles (green = pixels with G > R+12 and G > B+12 and G > 40; > 30 % of the tile counts as green-dominant).

Selection stops when no candidate satisfies all constraints. The final set: 136 tiles from 136 distinct photos.

6. Score sheet

val_scores.csv, one row per tile:

tile,source_id,category,green_fraction,complexity,entropy,lap_var,grad_energy,blockiness,aliasing,clipped,topiq_nr,clipiqa+,niqe,qrealign
000001.png,test_1079,animal,0.004,0.5625,7.602,1974.7,14.48,1.003,0.370,0.000,0.7711,0.8079,2.5280,0.8927

• complexity — calibrated ICNet score (higher = more detail)
• niqe — lower is better; all other scores higher is better
• category — CLIP zero-shot label (15-way), for per-content analysis
• green_fraction — the vegetation-dominance measure above

---

7. Usage

import webdataset as wds
gt = wds.WebDataset("gt/shard-00000.tar").decode("pil")
x4 = wds.WebDataset("x4/shard-00000.tar").decode("pil")
for (g,), (l,) in zip(gt, x4):
    # g: 512×512 GT (PIL), l: 128×128 LR (PIL)
    ...

• Validate every N training iterations on all 136 tiles — it is small enough to be effectively free, large enough for a stable curve.
• Report PSNR/SSIM/DISTS for x2 → gt and x4 → gt, bicubic as baseline.
• With 136 paired samples, treat differences below ~0.05 dB as noise.

8. Honest limits

• GT is one bicubic resample (4K → short side 512), not native pixels — the standard whole-scene convention, but absolute PSNR is not comparable to native-crop benchmarks.
• 136 scenes is comparable to Urban100's 100 images, but tiles are 512² crops, not full images — not directly comparable to published Urban100 numbers either. This set is an internal benchmark: fixed, reproducible, paired.
• The source corpus is aesthetic-curated and web-compressed; the gates remove the degradation classes we can measure, not every possible artifact.
• Content spread: interior/animal-heavy (the corpora are; per-category metrics can be computed from the score sheet's category column).

9. Provenance

• Tiles produced and filtered with the LUCID pipeline (repo); every threshold and formula is listed inline in this card, so the set is reproducible without the code.
• Selection script: DINOv2 (timm vit_base_patch14_dinov2.lvd142m) + constrained farthest-point as described in §5.
• Source imagery: UHD-IQA (CC0). Set license: CC0 1.0.
• Built by Philip Hofmann (BHI, BHI100, PSISRD).