PIM
Official repository for PIM checkpoints, evaluated on AI-generated image detection and segmentation tasks.
📌 Available Versions
| Version |
Best Metric Score |
Best Epoch |
Updated |
v1 |
0.852 |
10 |
2026-11-13 |
- Active Version:
v1
- Best Validation Score:
0.852 (Epoch 10)
- Total Training Epochs:
10
📊 Master Cross-Evaluation Results
SID-UNet Master Cross-Evaluation Report
Overview
- Evaluated Checkpoints: 2
- Evaluated Configurations: 1
- Total Cross-Evaluation Runs: 2
Cross-Evaluation Matrices
Mean IoU (Intersection over Union) Cross-Evaluation Matrix
| Checkpoint \ Config |
eval_mvss_v1_protocol |
| sid_focal_hrnet |
0.2293 |
| sid_pim_swin |
0.1645 |
Dice / Pixel F1 Score Cross-Evaluation Matrix
| Checkpoint \ Config |
eval_mvss_v1_protocol |
| sid_focal_hrnet |
0.2902 |
| sid_pim_swin |
0.1983 |
AUROC (Pixel-level ROC-AUC) Cross-Evaluation Matrix
| Checkpoint \ Config |
eval_mvss_v1_protocol |
| sid_focal_hrnet |
0.6171 |
| sid_pim_swin |
0.5770 |
Pixel Accuracy Cross-Evaluation Matrix
| Checkpoint \ Config |
eval_mvss_v1_protocol |
| sid_focal_hrnet |
0.7758 |
| sid_pim_swin |
0.8472 |
Evaluation Total Loss Cross-Evaluation Matrix
| Checkpoint \ Config |
eval_mvss_v1_protocol |
| sid_focal_hrnet |
1.0407 |
| sid_pim_swin |
0.9506 |
Master Consolidated Ranking
| Checkpoint |
Config (Dataset) |
Split |
Samples |
Loss |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Precision |
Recall |
Specificity |
Class Acc |
| sid_focal_hrnet |
eval_mvss_v1_protocol (KhangTruong/Tampered-Images-Combined-Dataset-Fake) |
cross_test |
17929 |
1.0407 |
0.2293 |
0.2902 |
0.6171 |
0.7758 |
0.2809 |
0.4618 |
0.8233 |
N/A |
| sid_pim_swin |
eval_mvss_v1_protocol (KhangTruong/Tampered-Images-Combined-Dataset-Fake) |
cross_test |
17929 |
0.9506 |
0.1645 |
0.1983 |
0.5770 |
0.8472 |
0.2761 |
0.2281 |
0.9416 |
N/A |
🖼️ Cross-Evaluation Heatmaps & Visual Illustrations
Cross Eval Iou Heatmap
Cross Eval F1 Heatmap
Cross Eval Pixel Acc Heatmap
Cross Eval Auroc Heatmap
Sample Predictions Mvss-V1-Protocol
Per-Checkpoint Deep-Dive
Checkpoint: sid_focal_hrnet
- Path:
outputs/RUN/focal_hrnet/checkpoints/checkpoint_best.pt
| Config |
Dataset |
Loss |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Samples |
| eval_mvss_v1_protocol |
KhangTruong/Tampered-Images-Combined-Dataset-Fake |
1.0407 |
0.2293 |
0.2902 |
0.6171 |
0.7758 |
17929 |
Checkpoint: sid_pim_swin
- Path:
outputs/RUN/pim_swin/checkpoints/checkpoint_best.pt
| Config |
Dataset |
Loss |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Samples |
| eval_mvss_v1_protocol |
KhangTruong/Tampered-Images-Combined-Dataset-Fake |
0.9506 |
0.1645 |
0.1983 |
0.577 |
0.8472 |
17929 |
Per-Configuration / Dataset Deep-Dive
Configuration: eval_mvss_v1_protocol (Dataset: KhangTruong/Tampered-Images-Combined-Dataset-Fake)
- Config Path:
configs/cross-eval/MVSS-v1-protocol.yaml
| Checkpoint |
Loss |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Samples |
| sid_focal_hrnet |
1.0407 |
0.2293 |
0.2902 |
0.6171 |
0.7758 |
17929 |
| sid_pim_swin |
0.9506 |
0.1645 |
0.1983 |
0.577 |
0.8472 |
17929 |
Per-Dataset Breakdown: sid_focal_hrnet on eval_mvss_v1_protocol
| Sub-Dataset |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Precision |
Recall |
Specificity |
Samples |
| CASIA_V1 |
0.6365 |
0.7179 |
0.9073 |
0.9578 |
0.6672 |
0.8478 |
0.9668 |
920 |
| CASIA_V2 |
0.7391 |
0.8137 |
0.9655 |
0.9641 |
0.7506 |
0.9649 |
0.9661 |
1000 |
| COCO_Inpaint |
0.0459 |
0.0794 |
0.5328 |
0.8406 |
0.0745 |
0.1849 |
0.8808 |
2000 |
| CocoGlide |
0.3778 |
0.4764 |
0.6976 |
0.7474 |
0.4808 |
0.5857 |
0.8095 |
507 |
| Columbia_Gray |
0.9122 |
0.9343 |
0.9541 |
0.9585 |
0.9255 |
0.9450 |
0.9632 |
180 |
| Coverage |
0.3435 |
0.4326 |
0.7408 |
0.7996 |
0.3552 |
0.6645 |
0.8172 |
100 |
| HiFi_IFDL |
0.1741 |
0.2129 |
0.5018 |
0.5265 |
0.2115 |
0.3835 |
0.6202 |
2000 |
| IMD2020 |
0.0366 |
0.0616 |
0.5838 |
0.8220 |
0.0394 |
0.3394 |
0.8282 |
2000 |
| IMD2020_RealLife |
0.3906 |
0.4944 |
0.8783 |
0.8673 |
0.4074 |
0.8846 |
0.8721 |
1000 |
| NIST16 |
0.1295 |
0.1965 |
0.2725 |
0.1882 |
0.5382 |
0.1385 |
0.4065 |
220 |
| OpenSDID_Flux.1 |
0.2272 |
0.3140 |
0.6525 |
0.7518 |
0.3027 |
0.4964 |
0.8087 |
2000 |
| OpenSDID_SD1.5 |
0.1420 |
0.2096 |
0.5808 |
0.7231 |
0.2200 |
0.3505 |
0.8112 |
2000 |
| OpenSDID_SDXL |
0.1309 |
0.1982 |
0.5655 |
0.7154 |
0.2187 |
0.3235 |
0.8075 |
2000 |
| SAGI-D |
0.5571 |
0.6475 |
0.8580 |
0.8672 |
0.5718 |
0.8130 |
0.9030 |
16 |
| TGIF2_flux1dev |
0.0030 |
0.0030 |
0.4422 |
0.8814 |
0.0030 |
0.0030 |
0.8814 |
1986 |
Per-Dataset Breakdown: sid_pim_swin on eval_mvss_v1_protocol
| Sub-Dataset |
IoU |
Dice / F1 |
AUROC |
Pixel Acc |
Precision |
Recall |
Specificity |
Samples |
| CASIA_V1 |
0.4965 |
0.5494 |
0.7653 |
0.9556 |
0.6277 |
0.5438 |
0.9868 |
920 |
| CASIA_V2 |
0.8464 |
0.8805 |
0.9123 |
0.9343 |
0.8829 |
0.8868 |
0.9378 |
1000 |
| COCO_Inpaint |
0.0075 |
0.0128 |
0.4928 |
0.9144 |
0.0247 |
0.0196 |
0.9659 |
2000 |
| CocoGlide |
0.2735 |
0.3429 |
0.6407 |
0.8141 |
0.5531 |
0.3075 |
0.9739 |
507 |
| Columbia_Gray |
0.5825 |
0.6561 |
0.7861 |
0.8614 |
0.7922 |
0.6195 |
0.9527 |
180 |
| Coverage |
0.2054 |
0.2665 |
0.6239 |
0.8878 |
0.3719 |
0.2776 |
0.9701 |
100 |
| HiFi_IFDL |
0.0959 |
0.1399 |
0.5702 |
0.6096 |
0.2182 |
0.3153 |
0.8250 |
2000 |
| IMD2020 |
0.0144 |
0.0235 |
0.5103 |
0.9496 |
0.0258 |
0.0593 |
0.9612 |
2000 |
| IMD2020_RealLife |
0.3364 |
0.4135 |
0.7185 |
0.9219 |
0.4713 |
0.4722 |
0.9647 |
1000 |
| NIST16 |
0.0380 |
0.0629 |
0.3727 |
0.1545 |
0.4306 |
0.0412 |
0.7041 |
220 |
| OpenSDID_Flux.1 |
0.0831 |
0.1187 |
0.5375 |
0.8234 |
0.2482 |
0.1162 |
0.9589 |
2000 |
| OpenSDID_SD1.5 |
0.0791 |
0.1111 |
0.5336 |
0.8172 |
0.2173 |
0.1115 |
0.9557 |
2000 |
| OpenSDID_SDXL |
0.0725 |
0.1054 |
0.5322 |
0.8115 |
0.2187 |
0.1117 |
0.9527 |
2000 |
| SAGI-D |
0.3337 |
0.3871 |
0.6746 |
0.8502 |
0.5611 |
0.3626 |
0.9867 |
16 |
| TGIF2_flux1dev |
0.0876 |
0.0876 |
0.5219 |
0.9562 |
0.0876 |
0.0876 |
0.9562 |
1986 |
🚀 Quickstart & Checkpoint Usage
1. Pull checkpoints using sid-pull
To download the latest checkpoint directly to your local workspace:
sid-pull --name KhangTruong/PIM
To pull a specific version:
sid-pull --name KhangTruong/PIM --version v1
2. Resume or Fine-tune with sid-train
Once downloaded, resume training seamlessly:
sid-train --config configs/experiments/diffusion_diff_minimized/default.yaml --override training.epochs=40
Or specify explicit checkpoint path:
sid-train --resume checkpoints/checkpoint_latest.pt
3. Push new checkpoints using sid-push
sid-push --name KhangTruong/PIM --version v2
📂 Repository Layout
KhangTruong/PIM/
├── README.md # Model Card & Documentation
├── config.yaml # Model & Training Configuration
├── manifest.json # Catalog of all versions & metrics
├── checkpoints/ # Active / Latest Checkpoints
│ ├── checkpoint_best.pt
│ ├── checkpoint_latest.pt
│ └── checkpoint_periodic.pt
├── reports/ # Evaluation reports & plots
│ ├── training_curves.png
│ └── test_evaluation_report.json
├── cross_evaluation/ # Master evaluation matrices & illustrations
│ ├── master_cross_evaluation_report.md
│ ├── cross_eval_matrix.json
│ └── illustrations/
└── versions/ # Version Archive
└── v1/
├── checkpoints/
├── reports/
└── cross_evaluation/
