diff --git a/deepguard/MS_EffGCViT.md b/deepguard/MS_EffGCViT.md index 2a5aa3b..7c4f5bf 100644 --- a/deepguard/MS_EffGCViT.md +++ b/deepguard/MS_EffGCViT.md @@ -14,7 +14,8 @@ This Repository presents the PyTorch implementation of **Multi Scale Efficient G This model is a **frame-level** and **spatial-domain** architecture, designed to perform classification tasks on both **static images** and **video sequences** - + + ## 💥 News 💥 @@ -27,21 +28,49 @@ This model is a **frame-level** and **spatial-domain** architecture, designed to ## Model Performance -MS_Eff_GCViT achieves state-of-the-art(SOTA) results across deepfake video classification. On Celeb_DF(v2) dataset, MS_EFF_GCViT variants with `8.7M`, `50.3M` parameters achieve `0.9842`, `0.9981` Accuracy. Notably, the MS_EFF_GCViT_B0 variant demonstrates exceptional efficiency, matching or exceeding SOTA performance even with a siginificantly lower parameter +**MS-EFF-GCViT achieves state-of-the-art (SOTA) results across three DeepFake benchmarks.** +The model ships in two variants from a single architecture — **Fast (b0)** for real-time / edge +deployment and **Pro (b5)** for enterprise-grade accuracy. Notably, **Fast** matches or exceeds +much larger SOTA models while using a fraction of the parameters and compute. +

+ +

-### Test Result of Celeb_DF(v2) +> On **Celeb-DF(v2)**, Pro reaches **0.9981 Acc** (rank #1) and Fast **0.9842** (rank #3) among 20 architectures. +> On the **KoDF competition** leaderboard, Pro ranks **#1** and Fast **#4** out of 49 entries. - +
+📊 Celeb-DF (v2) — Accuracy & Efficiency +
+ + +| Variant | Test@Acc | Test@AUC | Test@LogLoss | +| :------ | :------: | :------: | :----------: | +| ms_eff_gcvit_b0 | 0.9842 | 0.9965 | 0.0283 | +| ms_eff_gcvit_b5 | 0.9981 | 0.9984 | 0.0089 | +
-Test Result of FaceForensics++ - +📊 FaceForensics++ — Accuracy & Efficiency +
+ + +| Variant | Test@Acc | Test@AUC | Test@LogLoss | +| :------ | :------: | :------: | :----------: | +| ms_eff_gcvit_b0 | 0.9808 | 0.9969 | 0.0637 | +| ms_eff_gcvit_b5 | 0.9850 | 0.9974 | 0.0492 |
-Test Result of KoDF - +📊 KoDF Competition — Accuracy Ranking +
+ + +| Variant | Test@Acc | Test@AUC | Test@LogLoss | +| :------ | :------: | :------: | :----------: | +| ms_eff_gcvit_b0 | 0.9655 | 0.9792 | 0.1237 | +| ms_eff_gcvit_b5 | 0.9792 | 0.9974 | 0.0492 |
## Model Indroduction @@ -77,6 +106,8 @@ While both **Xception** and **EfficientNet** show great results on DeepFake benc ### Part 4: Multi-Scale Feature Map Fusion + + Modern DeepFakes can leave very localized forgery region. To Capture this, we adopts a **multi-scale strategy** by extracting features from different levels of the backbone. - **![](https://img.shields.io/badge/Low_level_Branch-blue?style=flat-square) (_Subtle Artifacts_)**: High-Resolution feature maps are extracted from early backbone blocks(`l_block_idx`) to capture like skin texture or boundary artifacts @@ -177,16 +208,6 @@ model = timm.create_model("ms_eff_gcvit_b5", pretrained=True, dataset="kodf") ## 📊 Visual Results -

- - - - - - -
-

- ### MS-EFF-GCVIT — Low-Level Branch | Model | Branch-Level | Image | HiresCam | GradCamElementwise | LayerCam | diff --git a/docs/benchmarks/celeb_df_v2_gcvit.png b/docs/benchmarks/celeb_df_v2_gcvit.png index 3704f61..7719047 100644 Binary files a/docs/benchmarks/celeb_df_v2_gcvit.png and b/docs/benchmarks/celeb_df_v2_gcvit.png differ diff --git a/docs/benchmarks/celeb_df_v2_gcvit_2.png b/docs/benchmarks/celeb_df_v2_gcvit_2.png new file mode 100644 index 0000000..5cc437f Binary files /dev/null and b/docs/benchmarks/celeb_df_v2_gcvit_2.png differ diff --git a/docs/benchmarks/ff_gcvit.png b/docs/benchmarks/ff_gcvit.png index 536e068..9a70a73 100644 Binary files a/docs/benchmarks/ff_gcvit.png and b/docs/benchmarks/ff_gcvit.png differ diff --git a/docs/benchmarks/gcvit_summary_bars.png b/docs/benchmarks/gcvit_summary_bars.png new file mode 100644 index 0000000..ee8d30b Binary files /dev/null and b/docs/benchmarks/gcvit_summary_bars.png differ diff --git a/docs/benchmarks/kodf_gcvit.png b/docs/benchmarks/kodf_gcvit.png index f90a973..d163b38 100644 Binary files a/docs/benchmarks/kodf_gcvit.png and b/docs/benchmarks/kodf_gcvit.png differ