sawabedarain/lumina-iqa

Lumina v2 — perceptual image quality with spatial explanations

No-reference image quality assessment (NR-IQA). Given a photograph, Lumina predicts the mean opinion score (0–100, KonIQ-10k scale) a panel of human viewers would give it, and explains the score:

• quality map: local quality for every 16×16 region
• attention map: where viewers look; the score is the attention-weighted mean of the quality map
• six defect maps + scores: blur, noise, artifacts, exposure, contrast, colour
• five perceptual attributes (SPAQ-calibrated): sharpness, noise-free, contrast, colourfulness, brightness

Architecture

CLIP-pretrained ConvNeXt-Base (convnext_base.clip_laion2b_augreg_ft_in12k_in1k_384), stride-8/16/32 features fused on the 1/16 grid, four light heads. 92M parameters. Fully convolutional: images are scored at native aspect (long side ≤ 512 px, short ≤ 384 px).

Training

Joint multi-source training on an RTX 5090 (bf16, EMA, AdamW, cosine schedule, 40 epochs): KonIQ-10k MOS (official 7,058 train split) · KonIQ++ defect votes · SPAQ MOS + attributes (8,900) · KADID-10k MOS + distortion families (8,190, reference-disjoint split) · on-the-fly synthetic degradations with a self-supervised ranking constraint (degraded copy must score below its original). Losses: smooth-L1 + PLCC + pairwise rank on MOS, masked BCE on defects, L1 on attributes; per-dataset affine calibration.

Results

Test set	v2 SRCC	v2 PLCC	v1 SRCC	v1 PLCC
KonIQ-10k (official test, 2,015)	0.940	0.950	0.894	0.919
LIVE Challenge (zero-shot, 1,169)	0.830	0.837	0.742	0.776
SPAQ (held-out 20%, 2,225)	0.930	0.934	0.851	0.843
KADID-10k (held-out reference scenes)	0.882	0.887	0.521	0.543

v1 = the original EfficientNet-B0 regressor. It was trained on a random split of all of KonIQ, so its KonIQ test number is inflated by train/test overlap; LIVE-C and SPAQ are the fair comparisons. v2 numbers use horizontal-flip TTA.

Files

• lumina_v2.int8.onnx: dynamic int8 (MatMul) for CPU serving
• lumina_v2.onnx: fp32
• lumina_v2.json: normalisation, KonIQ score quantiles (for percentiles), label names
• metrics.json, history.json: evaluation and per-epoch validation

Input: image float32 RGB in [0, 255], shape (N, 3, H, W). Outputs: mos, qmap, wmap, defects, dmap, attrs.

Source & training code: https://github.com/DarainHyder/Image_Quality_Assessment