86Cao/IQA-ONNX-Models

IQA ONNX Models

Pre-converted ONNX models for No-Reference Image Quality Assessment (NR-IQA), exported from IQA-PyTorch and ready for deployment with ONNX Runtime (no PyTorch dependency needed).

Models

Model	Input Size	Files	Total Size	Score Range	Description
LIQE	224×224	clip_model.onnx + liqe_model.onnx + text_features.json	~675 MB	1–5	CLIP-based learned quality & scene & distortion assessment
DBCNN	224×224	dbcnn_model.onnx	~59 MB	0–100	Double-blind CNN with VGG16 + SCNN bilinear pooling
HyperIQA	224×224	hyperiqa_model.onnx	~105 MB	0–1	Hyper-network based quality prediction on ResNet50
MANIQA	224×224	maniqa_model.onnx	~437 MB	continuous	Multi-dimension Attention Network with ViT backbone
MUSIQ	224×224	musiq_model.onnx	~104 MB	0–100	Multi-scale Image Quality Transformer (single-scale export)
TReS	224×224	tres_model.onnx	~575 MB	0–100	Transformer + ResNet ensemble with relative ranking
CLIPIQA	224×224	clipiqa_model.onnx	~146 MB	0–1	CLIP-IQA+ with learned prompts and antialias RN50

Most models use external data format (.onnx + .onnx.data). Both files must be in the same directory for loading.

Quick Start

import onnxruntime as ort
import numpy as np
from PIL import Image

# Load model
sess = ort.InferenceSession("dbcnn_model.onnx")

# Preprocess: resize to 224x224, normalize with ImageNet mean/std
img = Image.open("test.jpg").convert("RGB").resize((224, 224))
x = np.array(img).astype(np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
x = (x - mean) / std
x = x.transpose(2, 0, 1)[np.newaxis]  # (1, 3, 224, 224)

# Run inference
score = sess.run(None, {"input": x.astype(np.float32)})[0]
print(f"Quality score: {score.item():.4f}")

Preprocessing

Model	Normalization	Notes
DBCNN, HyperIQA, TReS	ImageNet: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]	Standard ImageNet
MANIQA	Inception: mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]
MUSIQ	[-1, 1]: (x - 0.5) / 0.5	Equivalent to Inception
CLIPIQA	CLIP: mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711]
LIQE	CLIP: same as CLIPIQA	Two-stage: CLIP encoder → LIQE scorer

For multi-crop inference (recommended for HyperIQA, MANIQA, TReS), extract multiple 224×224 crops from the full-resolution image and average the scores.

LIQE Usage

LIQE requires a two-stage pipeline:

import json
import onnxruntime as ort
import numpy as np

# Load both models and text features
clip_sess = ort.InferenceSession("clip_model.onnx")
liqe_sess = ort.InferenceSession("liqe_model.onnx")
with open("text_features.json") as f:
    text_features = np.array(json.load(f), dtype=np.float32)

# Step 1: Extract image features with CLIP
img_input = preprocess_clip(image)  # (1, 3, 224, 224)
img_features = clip_sess.run(None, {"input": img_input})[0]

# Step 2: Score with LIQE
score = liqe_sess.run(None, {
    "image_features": img_features,
    "text_features": text_features
})[0]

Verification Results

All models verified against PyTorch source (absolute difference < 0.01):

Model	PyTorch Score	ONNX Score	Abs Diff
LIQE	—	—	< 0.001
DBCNN	—	—	0.000011
HyperIQA	—	—	0.00000003
MANIQA	—	—	0.000019
MUSIQ	—	—	0.000025
TReS	—	—	0.000439
CLIPIQA	—	—	0.000002

Export Details

• Source: IQA-PyTorch v0.1.15
• PyTorch version: 2.10.0+cu128
• ONNX opset: 14–18 (varies by model)
• Pretrained on: KonIQ-10k (all models)
• Export mode: eval-only forward pass (no training-specific branches)

Per-Model Conversion Details

Model	Paper	Conference	ONNX Files	Export Notes
LIQE	Blind Image Quality Assessment via Vision-Language Correspondence	CVPR 2023	clip_model.onnx liqe_model.onnx text_features.json	Two-stage: CLIP ViT-B/32 image encoder + LIQE scoring head; text features pre-encoded to JSON
DBCNN	Blind Image Quality Assessment Using A Deep Bilinear CNN	IEEE TCSVT 2020	dbcnn_model.onnx	VGG16 + SCNN with bilinear pooling; both sub-networks exported as single model
HyperIQA	Blindly Assess Image Quality in the Wild Boosted by A Large-scale Database	CVPR 2020	hyperiqa_model.onnx	ResNet50 backbone with hyper-network; exported forward patch path only
MANIQA	MANIQA: Multi-dimension Attention Network for No-Reference Image Quality Assessment	CVPR 2022 NTIRE	maniqa_model.onnx	ViT-B/8 backbone; single-file export (no external data)
MUSIQ	MUSIQ: Multi-scale Image Quality Transformer	ICCV 2021	musiq_model.onnx	Simplified to single-scale 224×224 input (original uses multi-scale patches)
TReS	No-Reference Image Quality Assessment via Transformers, Relative Ranking, and Self-Consistency	WACV 2022	tres_model.onnx	Eval-only path; dual-path consistency and flip branches removed
CLIPIQA	Exploring CLIP for Assessing the Look and Feel of Images	AAAI 2023	clipiqa_model.onnx	CLIP-IQA+ variant with learned prompts baked into the model; antialias RN50 backbone

File List

clip_model.onnx           # LIQE CLIP image encoder
clip_model.onnx.data
liqe_model.onnx           # LIQE scoring head
liqe_model.onnx.data
text_features.json        # LIQE pre-encoded text prompts
dbcnn_model.onnx          # DBCNN
dbcnn_model.onnx.data
hyperiqa_model.onnx       # HyperIQA
hyperiqa_model.onnx.data
maniqa_model.onnx         # MANIQA (single file, no .data)
musiq_model.onnx          # MUSIQ
musiq_model.onnx.data
tres_model.onnx           # TReS
tres_model.onnx.data
clipiqa_model.onnx        # CLIPIQA+
clipiqa_model.onnx.data

Citation

If you use these models, please cite the original papers and the IQA-PyTorch toolbox:

@article{chaofeng2022iqapytorch,
  title={IQA-PyTorch: PyTorch Toolbox for Image Quality Assessment},
  author={Chaofeng Chen and Jiadi Mo},
  year={2022},
  journal={arXiv preprint arXiv:2208.14818}
}

License

Apache 2.0. Original model weights are subject to their respective licenses.