SigLIP2 NR-IQA MLX
Collection
MLX no-reference image-quality head on SigLIP2-SO400M (repro of arXiv:2509.17374). • 1 item • Updated • 1
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 — 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
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.
- Downloads last month
- 47
Hardware compatibility
Log In to add your hardware
Quantized
Model tree for mlx-community/SigLIP2-NR-IQA-KonIQ
Base model
google/siglip2-so400m-patch16-512