Paper (PDF) | arXiv | WACV 2026 Open Access

Ankit Yadav, Ta Duc Huy, Lingqiao Liu The University of Adelaide

We present the first systematic evaluation of six prominent pretrained backbones (CLIP, SigLIP2, DINOv2, DINOv3, Perception, and ResNet) for No-Reference Image Quality Assessment (NR-IQA). Our study uncovers that (1) SigLIP2 consistently achieves strong performance, and (2) the choice of activation function plays a surprisingly crucial role. We introduce a learnable activation selection mechanism that adaptively determines the nonlinearity for each channel, achieving new state-of-the-art SRCC on CLIVE, KADID10K, and AGIQA3K.

git clone https://github.com/drkkgy/NR_IQA_AGM.git && cd NR_IQA_AGM
pip install -r requirements.txt

# Reproduce every paper number in one command (skips runs whose dataset is missing)
python eval_all.py

# Discover the available pretrained runs, then evaluate one by name
python eval_checkpoint.py --list
python eval_checkpoint.py --run Gating_CLIVE

# Evaluate the B_Gated (gating) checkpoint with GradCAM heatmaps
python eval.py --dataset CLIVE

Loss: MSE + pair-wise margin ranking loss

Create a Dataset/ folder in the project root and organise each benchmark as shown below. The exact sub-directory names and annotation files must match what dataset.py expects.

Dataset/
├── KonIQ_10K/
│   ├── koniq10k_512x384/
│   │   └── 512x384/                # 10,073 images
│   └── koniq10k_scores_and_distributions/
│       └── koniq10k_scores_and_distributions.csv
│
├── CLIVE/
│   └── ChallengeDB_release/
│       ├── Data/
│       │   ├── AllImages_release.mat
│       │   ├── AllMOS_release.mat
│       │   └── AllStdDev_release.mat
│       └── Images/                  # 1,162 images
│
├── SPAQ/
│   ├── SPAQ_dataset/
│   │   └── Annotations/
│   │       └── MOS_and_Image_attribute_scores.xlsx
│   └── TestImage/                   # 11,125 images
│
├── KADID-10K/
│   └── kadid10k/
│       ├── dmos.csv
│       └── images/                  # 10,125 images
│
├── FLIVE/
│   ├── labels_image.csv
│   └── database/                    # ~40,000 images (sub-folders inside)
│
├── AGIQA-3k/
│   ├── data.csv
│   └── images/                      # 2,982 images
│
└── AGIQA-1k/
    ├── AIGC_MOS_Zscore.xlsx
    └── images/                      # 1,000 images

Tip: You can symlink existing dataset directories instead of copying:

ln -s /path/to/your/KonIQ_10K Dataset/KonIQ_10K

pip install -r requirements.txt

Note: Install PyTorch with the CUDA version matching your GPU driver first. See pytorch.org/get-started.

conda env create -f environment.yml
conda activate nr_iqa_agm

Edit pytorch-cuda=12.1 in environment.yml if you need a different CUDA version (e.g. 11.8).

# Train on KonIQ-10K with default hyperparameters
python train.py --dataset KonIQ_10K

# Train on CLIVE with LoRA rank 8, 20 epochs, batch size 4
python train.py --dataset CLIVE --peft_method LoRA --lora_r 8 --epochs 20 --batch_size 4

# Cross-dataset: train on KonIQ-10K, evaluate on CLIVE
python train.py --dataset KonIQ_10K_CLIVE

# Full fine-tuning (no PEFT adapter)
python train.py --dataset SPAQ --peft_method NA

# Deep Prompt Tuning instead of LoRA
python train.py --dataset KADID10K --peft_method DPT

# Resume a previous run
python train.py --dataset KonIQ_10K --resume

# Dry-run for quick debugging (100 train batches, 32 eval batches)
python train.py --dataset CLIVE --dry_run

# Disable WandB logging
python train.py --dataset CLIVE --no_wandb

The repo ships with seven pretrained runs in pretrained_checkpoints/, covering two architectures (B_Gated with MLP3_Gated activation gating, and B_Sig with mlp_3_layer_sigmoid_siglip), each as a LoRA adapter on SigLIP-2. Reference them by their short run name:

Each registered run carries the same train/val split and seed used during training, so the reported SRCC/PLCC on the held-out partition match the paper. Pass --run <name> to eval_checkpoint.py and the right split is applied for you — no extra flags needed.

Run python eval_checkpoint.py --list to print the same table with the underlying checkpoint paths. eval_all.py iterates every entry and prints an SRCC/PLCC summary — see the Evaluation section below.

Performance comparison with state-of-the-art methods on seven benchmark datasets. Values represent SRCC and PLCC averaged over three runs (seeds: 8, 19, 25). B: Baseline, B_Sig: Baseline_Sigmoid, B_Gated: Baseline_Gated (Ours).

There are two evaluation scripts:

• eval.py — evaluates B_Gated (MLP3_Gated) checkpoints with optional GradCAM visualisation.
• eval_checkpoint.py — unified script that supports all three architectures (Baseline, B_Sig, B_Gated) with auto-detection.

# Evaluate using the pretrained CLIVE->CLIVE checkpoint (auto-detected)
python eval.py --dataset CLIVE

# Cross-dataset: pretrained KonIQ_10K->CLIVE checkpoint
python eval.py --dataset KonIQ_10K_CLIVE

# Evaluate a specific (user-trained) checkpoint
python eval.py --dataset KonIQ_10K \
    --checkpoint_dir best_checkpoints/AGM_seed8_train_KonIQ_10K_test_KonIQ_10K

# Skip GradCAM visualisation
python eval.py --dataset SPAQ --no_gradcam

# Custom output path
python eval.py --dataset AGIQA3K --output my_results.json

Unified script supporting Baseline, B_Sig, and B_Gated architectures. Architecture is auto-detected from the checkpoint's mlp.pt state-dict keys (and the directory name) — manual override is available via --arch.

All shipped checkpoints are PEFT/LoRA adapters on top of google/siglip2-so400m-patch16-512, loaded with peft.PeftModel.from_pretrained() and merged for clean inference.

# List all pretrained runs
python eval_checkpoint.py --list

# Evaluate one by name (resolves checkpoint + arch + dataset from the registry)
python eval_checkpoint.py --run Gating_CLIVE
python eval_checkpoint.py --run Sigmoid_KonIQ_to_CLIVE

# Override the registry's test dataset (cross-dataset experiment with the same checkpoint)
python eval_checkpoint.py --run Gating_KonIQ --dataset AGIQA3K

# Pick a specific GPU and save to a custom location
python eval_checkpoint.py --run Gating_CLIVE --device cuda:1 --output clive.json

For checkpoints not in the registry, evaluation uses the same 20% held-out split as training — pass --seed so the partition can be reconstructed deterministically (use the same seed you trained with).

# Evaluate a user-trained checkpoint on its held-out 20% partition
python eval_checkpoint.py \
    --checkpoint best_checkpoints/AGM_seed8_train_CLIVE_test_CLIVE \
    --dataset CLIVE \
    --seed 8

# Manually override the architecture if auto-detection picks wrong
python eval_checkpoint.py \
    --checkpoint /path/to/baseline_sigmoid_ckpt \
    --dataset KonIQ_10K \
    --seed 8 \
    --arch baseline_sig

# Run every pretrained checkpoint on its native test set
python eval_all.py

# Run only the gating-architecture checkpoints
python eval_all.py --filter Gating

# Larger batch size, alternate dataset root, save to a custom location
python eval_all.py --data_dir /data/IQA --batch_size 8 --output my_sweep.json

# Fail loudly on the first missing dataset (default: skip-with-warning)
python eval_all.py --strict

Runs whose dataset is not present under --data_dir are skipped with a warning by default — useful when you only have a subset of benchmarks downloaded. After the sweep, an SRCC/PLCC summary table is printed to stdout and combined results are saved to results/eval_all.json.

• SRCC (Spearman Rank-order Correlation Coefficient): measures monotonic association between predicted and ground-truth scores.
• PLCC (Pearson Linear Correlation Coefficient): measures linear correlation after fitting.

Pass a combined dataset ID to train.py:

• KonIQ_10K_CLIVE — train on KonIQ-10K, evaluate on CLIVE
• CLIVE_KonIQ_10K — train on CLIVE, evaluate on KonIQ-10K

See LICENSE.

• Add pretrained checkpoints for KonIQ-10K (B_Gated, B_Sig)
• Add cross-dataset pretrained checkpoints (KonIQ-10K -> CLIVE, CLIVE -> KonIQ-10K)
• Add Sigmoid_CLIVE_to_KonIQ checkpoint (B_Sig trained on CLIVE, tested on KonIQ-10K)
• Add pretrained checkpoints for SPAQ, KADID-10K, FLIVE, AGIQA-3K, AGIQA-1K

If you use this codebase in your research, please cite:

@InProceedings{Yadav_2026_WACV,
    author    = {Yadav, Ankit and Huy, Ta Duc and Liu, Lingqiao},
    title     = {Revisiting Vision-Language Foundations for No-Reference Image Quality Assessment},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {March},
    year      = {2026},
    pages     = {5416-5425}
}