RobustDeepfakeCVPR26 — AI4Good greedy6 ensemble
Model checkpoints for Team AI4Good / SoloUNI's submission to the NTIRE 2026 Robust Deepfake Detection Challenge (CVPR 2026 Workshop).
Method — Self-Supervised Adversarial Training. Binary classifiers trained on the 500 real images only using Self-Blended Images (SBI) + PMM degradation. No real fakes are used — pseudo-fakes are generated on-the-fly from pairs of real images via landmark-guided Poisson/Alpha blending. Six diverse backbones are trained independently and combined by probability averaging with 3× deterministic TTA.
- Code & reproduction: https://github.com/CristianLazoQuispe/RobustDeepfakeCVPR26
- Method writeup (arXiv): https://arxiv.org/abs/2604.24163
- Challenge report (CVPR 2026): Hopf et al., Robust Deepfake Detection, NTIRE 2026 Challenge: Report
- License: MIT
The 6 models
Each file is <model>/best.pt — a PyTorch checkpoint dict with keys including model (state_dict),
epoch, best_auc, and the full training config. All trained at 224×224, AdamW, cosine 100 epochs,
seed: 42 (csl uses 43), deterministic: true, amp: false.
| Model | Backbone (timm) | In ch. | Hard-Test AUC | Val AUC |
|---|---|---|---|---|
r3d_fa_csb | convnext_base.fb_in22k_ft_in1k | 3 | 0.7644 | 0.897 |
lmhm_css | convnext_small.fb_in22k_ft_in1k + LMHM | 4 | 0.7268 | 0.900 |
r3e_fa_css | convnext_small.fb_in22k_ft_in1k | 3 | 0.7259 | 0.844 |
r3d_fa_csl_s43 | convnext_large.fb_in22k_ft_in1k | 3 | 0.7127 | 0.867 |
r3d_fa_deit3_s | deit3_small_patch16_224 | 3 | 0.7111 | 0.889 |
lmhm_vits | vit_small_patch16_224.augreg_in21k_ft_in1k + LMHM | 4 | 0.6945 | 0.899 |
| Ensemble (6) + TTA | — | — | 0.7777 | — |
CodaBench Public Test AUC: 0.6865 (submission 632762, 6-model ensemble + TTA). Hard-Test AUC is on a local 500-real/500-fake degraded holdout; Val AUC is the in-training best.
LMHM= landmark-heatmap models take a 4-channel input (RGB + face-landmark heatmap, σ=4).
Usage
from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="CristianLazoQuispe/RobustDeepfakeCVPR26-models", repo_type="model")
# place each <model>/best.pt under $DATA_ROOT/experiments/<model>/fold0/best.pt, then:
# Reproduce the submission with the code repo (6-model ensemble + TTA)
python scripts/generate_ensemble_submission.py \
--models r3d_fa_csb r3e_fa_css lmhm_vits r3d_fa_deit3_s r3d_fa_csl_s43 lmhm_css \
--device cuda:0 --tta --output greedy6_final.zip
Architecture & preprocessing for each model are embedded in its checkpoint's config. The two
lmhm_* models require the landmark-heatmap 4th channel (see the code repo's dataset/transform).
Citation
@InProceedings{Hopf_2026_CVPR,
author = {Hopf, Benedikt and Timofte, Radu and others},
title = {Robust Deepfake Detection, NTIRE 2026 Challenge: Report},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2026},
pages = {1925-1938}
}