init: project structure with LAION scorer, conda setup, and research docs

2026-08-23 15:31:47 +00:00
commit 4f90f5838a
25 changed files with 1628 additions and 0 deletions
@@ -0,0 +1,65 @@
+# IQA-PyTorch Toolbox — All-in-One
+
+## Model Info
+
+- **GitHub:** <https://github.com/chaofengc/IQA-PyTorch>
+- **PyPI:** `pip install pyiqa`
+- **License:** Apache 2.0
+- **Contents:** 30+ IQA metrics including MUSIQ, TOPIQ, BRISQUE, NIQE, LPIPS, FID, NIMA, DBCNN, CLIP-IQA, LIQE, Q-Align (Q-ReAlign), and more
+- **Latest updates (2026):**
+  - Jun 2026: Added Q-ReAlign (Qwen3.5-VL backbone) — 3 sizes: mini (0.8B), lite (4B), pro (9B)
+  - May 2026: Added FGResQ
+  - Dec 2025: Added DMM, MACLIP, AFINE
+  - Jan 2025: Added QualiCLIP variants
+- **Results calibrated:** Against official MATLAB scripts when available
+- **GPU accelerated:** PyTorch implementations are faster than MATLAB counterparts
+
+## Why It Matters
+
+This toolbox lets you test multiple IQA models on your wedding photos without installing each separately. Useful for benchmarking and comparison. It's the single easiest way to run BRISQUE + NIQE + MUSIQ + other metrics in one codebase.
+
+```bash
+# List all available metrics
+pyiqa -ls
+
+# Test multiple metrics on a directory
+pyiqa musiq niqe brisque -t ./wedding-photos/ --device cuda
+
+# Python API — all metrics in one import
+import pyiqa
+import torch
+
+device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+# Create any metric
+brisque = pyiqa.create_metric('brisque', device=device)
+musiq = pyiqa.create_metric('musiq', device=device)
+niqe = pyiqa.create_metric('niqe', device=device)
+
+# Check scoring direction
+print(f"BRISQUE lower_better: {brisque.lower_better}")  # True
+print(f"MUSIQ lower_better: {musiq.lower_better}")      # False (higher = better)
+print(f"NIQE lower_better: {niqe.lower_better}")        # True
+
+# Batch inference
+import glob
+for path in glob.glob('./wedding-photos/*.jpg'):
+    score_b = brisque(path)
+    score_m = musiq(path)
+    score_n = niqe(path)
+```
+
+## Available Metrics (relevant to wedding photo sorting)
+
+| Metric | Type | Score Direction | GPU Required | Notes |
+|--------|------|----------------|--------------|-------|
+| brisque | NR-IQA | Lower better | No (CPU) | Blur/noise detection |
+| niqe | NR-IQA | Lower better | No (CPU) | Blur/noise detection, opinion-unaware |
+| musiq | NR-IQA (Transformer) | Higher better | Yes (~4GB) | Aesthetic + technical combined |
+| toqip | NR-IQA (Transformer) | Higher better | Yes (~2GB) | Modern, self-supervised |
+| clip_iqa | NR-IQA (CLIP) | Higher better | Yes (~2GB) | CLIP-based aesthetic |
+| liqe | NR-IQA (CLIP) | Higher better | Yes (~2GB) | CLIP-based quality |
+| lpips | FR-IQA (Perceptual) | Lower better | Yes (~1GB) | Needs reference image |
+| fid | FR-IQA (Distribution) | Lower better | Yes (~5GB) | Needs reference dataset |
+| nima | NR-IQA (CNN) | Higher better | Yes (~1GB) | Neural IMage Assessment |
+| dbcnn | FR-IQA (CNN) | Higher better | Yes (~2GB) | Needs reference image |