Contest solution for distinguish AIGC images
Contest Page: https://xihe.mindspore.cn/competition/mindcon23-aigc-img/0/introduction
Team Name: 你这图保真吗
Submission repo: https://xihe.mindspore.cn/projects/kahsolt/AIGC-clf
Final Score/Rank: 71%/7th, no award :( (refer to SOLUTIONS.md for top solutions)
⚪ migrated pretrained apps
method | pAcc | comment |
---|---|---|
cheaty | 99.897224% | detect image h/w==512 |
AI-generated-art-classifier | 88.69476% | resnet18 clf |
AI-image-detector | 78.82837% | swin clf |
sdxl-detector | 56.62898% | swin clf |
sd-vae-ft-ema | 70.7086% | aekl + clf by loss_diff |
⚪ finetuned apps
ℹ following apps are based on AI-generated-art-classifier
and sd-vae-ft-ema
- run
python predict.py --app <app_name>
app | input_size | pAcc | pAcc (vote=5 ) |
pAcc (vote=7 ) |
comment |
---|---|---|---|---|---|
resnet | 224 | 85.71429% | 85.50874% | 85.09764% | migrated baseline |
resnet_ft | 224 | 98.86948% | 99.07503% | 99.07503% | finetune |
resnet_hf | 80 | 87.97533% | 93.01131% | 93.62795% | retrain from pretrained |
aekl_clf | 256 | 95.67456% | 95.88900% | 96.60843% | finetune from pretrained |
⚪ ensemble app
- run
python predict_ensemble.py --votes 7
app | pAcc | pAcc (vote=7 ) |
---|---|---|
resnet_ft | 96.40288% | 97.22508% |
resnet_hf | 89.31141% | 96.71120% |
aekl_clf | 96.50565% | 97.43063% |
ensembled | 98.15005% | 99.38335% |
⚪ run pretrained apps
- install PyTorch
pip install -r requirements.txt
- run
python predict.py --app <app_name>
and seeout/result.txt
⚪ finetune the apps
- link the contest dataset to
data
- run the following train scripts
python train_resnet_ft.py
python train_resnet_hf.py
python train_aekl_clf.py
- AI-detectors
- SD VAEs
- dataset
- 比赛材料仓库
by Armit 2024/01/03