SigLIP2 NR IQA KonIQ

SRCC 0.9575 · PLCC 0.9678 (paper reports SRCC ~0.932). Higher = better quality.

Usage

from siglip2_nriqa_mlx.score import NRIQAScorer
scorer = NRIQAScorer.from_pretrained("adapter.safetensors")  # backbone auto-loaded
print(scorer.score("photo.jpg"))   # ~0-100 MOS scale

Apache-2.0 (backbone Apache-2.0, Google). Training data KonIQ-10k is research-use — validate licensing for commercial deployment. See NOTICE.

--- license: apache-2.0 library_name: mlx pipeline_tag: image-classification tags: [mlx, image-quality-assessment, nr-iqa, siglip2, lora] base_model: google/siglip2-so400m-patch16-512 --- # SigLIP2 NR-IQA (KonIQ-10k) — MLX No-reference image-quality head on **SigLIP2-SO400M** for Apple Silicon (MLX). Independent from-paper reproduction of [arXiv:2509.17374](https://arxiv.org/abs/2509.17374) — no upstream code/weights were released. The first MLX no-reference IQA head. Ships the trained **LoRA(q/k) + 3-layer MLP head** adapter (~5 MB); the SigLIP2-SO400M backbone loads from the Google release at runtime. ## Results (KonIQ-10k, 80/20 val) **SRCC 0.9575 · PLCC 0.9678** (paper reports SRCC ~0.932). Higher = better quality. ## Usage ```python from siglip2_nriqa_mlx.score import NRIQAScorer scorer = NRIQAScorer.from_pretrained("adapter.safetensors") # backbone auto-loaded print(scorer.score("photo.jpg")) # ~0-100 MOS scale ``` Apache-2.0 (backbone Apache-2.0, Google). Training data KonIQ-10k is research-use — validate licensing for commercial deployment. See `NOTICE`.