diff --git a/AI_POLICY.md b/AI_POLICY.md
new file mode 100644
index 00000000..8556bf54
--- /dev/null
+++ b/AI_POLICY.md
@@ -0,0 +1,16 @@
+We support the use of AI tools to help prepare issues, pull requests, reviews or comments. We expect everyone interacting with this repo to follow the below policy whenever they use AI tools.
+
+AI-generated content in comments, issues, or PRs must be clearly disclosed and contained, either under a section heading that labels it as AI-generated or in a code or quote block. AI-generated content must be acoompanied by human commentary explaining ist relevance.  For example:
+"Codex produced the following analysis: `<insert codex output here>`, so I
+believe that \<insert human analysis here\>". The only exceptions to this rule
+are the dependabot automations.
+
+*For pull requests, please carefully read the code before submitting it for
+review*, especially if AI tools helped write or rewrite it. Make sure the
+implementation is something you understand, that the important ideas are clear,
+and that the code is high quality. AI-generated code can sometimes be overly
+complex, indirect, inconsistent, or include artifacts that obscure the main idea.
+Before submitting, simplify where possible and make sure the core change is easy
+for maintainers to review.
+
+All contributions should involve a human who understands and can take responsibility for the work produces with AI assistance. We do not accept contributions created by fully autonomous agents, and we may close pull request that appear to have been generated without meaningful human involvement.
\ No newline at end of file
diff --git a/CLAUDE.md b/CLAUDE.md
index fb593912..00aaf7a8 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -2,6 +2,26 @@
 
 This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
 
+# AI Policy — MANDATORY
+
+Read `AI_POLICY.md`. Your user needs to abide by this policy. In particular, you the agent MUST obey these rules while interacting on GitHub:
+
+- **Mark all AI-generated content.** Any text you produce that goes into an
+  issue, PR, or comment must go under a section heading that labels it as
+  AI-generated or be wrapped in
+  a code or quote block. Never present your output as human-written.
+- **Never emit only raw AI text as a reply**. Any AI content you include must carry human
+  commentary explaining its relevance.
+- **Do not submit code the user hasn't read.** Keep changes minimal, strip AI
+  artifacts and needless complexity. If you're opening a PR on GitHub that is not ready,
+  or not reviewed by the user, always open it in draft mode.
+- **Check for duplicates.** Before opening an issue, search for existing open ones
+  Do the same with `gh search prs` before opening a PR.
+  If you find a likely duplicate, show it to the user before continuing. Skip this
+  step if there is no internet access.
+
+See `AI_POLICY.md` for the full policy.
+
 ## Repository overview
 
 DeepGuard is a deepfake detection research project built around two custom hybrid CNN-ViT
diff --git a/README.md b/README.md
index 5334b301..1c53a18e 100644
--- a/README.md
+++ b/README.md
@@ -25,7 +25,6 @@
   <img src="https://img.shields.io/badge/Model-MS--Eff--GCViT--B0%20%2F%20B5-orange?style=flat-square" alt="Models">
   <img src="https://img.shields.io/badge/Python-3.10+-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python">
   <img src="https://img.shields.io/badge/PyTorch-EE4C2C?style=flat-square&logo=pytorch&logoColor=white" alt="PyTorch">
-  <img src="https://img.shields.io/badge/W%26B-Recording-FFBE00?style=flat-square&logo=weightsandbiases&logoColor=white" alt="W&B">
 </p>
 
 <p align="center">
@@ -44,12 +43,14 @@
 
 ## 📌 Contents
 
+- [🗂 Project Structure](#-project-structure) - Directory layout of the package, web app, and docs
 - [🐳 Docker Quick Start](#-docker-quick-start) - Run the full stack (MySQL, Redis, FastAPI, Celery, React) with Docker Compose
 - [🤗 Try It Live: Hugging Face Spaces](#-try-it-live-hugging-face-spaces) - No-install browser demos for image, video, and XAI deepfake detection
 - [📚 DeepFake Video BenchMark Datasets](#-deepfake-video-benchmark-datasets) — Overview of Celeb-DF-v2, FF++, and KoDF datasets used for training.
 - [⚙️ Data Preparation](#data-preparation) — Efficient face detection and landmark extraction pipeline using YOLOv8
 - [🏗 Model Architecture](#-model-architecture) — Detailed look into our hybrid CNN-ViT (MS-EffViT & MS-EffGCViT) designs.
 - [🧬 Model Zoo](#-model-zoo) — Comparison of model variants, parameter counts, and computational complexity (FLOPs).
+- [🕐 Model Inference Latency](#-model-inference-latency) - Fast (b0) vs Pro (b5) inference speed on a single Tesla T4
 - [🚀 Training](#-training) - Step-by-step training scrips with Goolge Colab and W&B experiment tracking
 - [📈 Model Evaluation](#-model-evaluation) - Benchmarking results
 - [💻 Model Usage](#-model-usage) - Load pretrained models via `pip install deepguard` or straight from the Hugging Face Hub
@@ -60,6 +61,29 @@
 - [📝 Reference](#-reference) - Libraries, datasets, and prior work this project builds on
 - [⚖️ License](#-license) - MIT license
 
+## 🗂 Project Structure
+
+```
+DeepGuard/
+├── deepguard/                  # Core package (models, layers, data pipeline, optimizer, scheduler, utils)
+├── preprocess/                 # Face detection, cropping & landmark extraction pipeline
+├── inference/                  # Image / video prediction
+├── explainability/             # Dual-branch CAM-based XAI
+├── App/                        # FastAPI backend (excluded from PyPI package)
+├── client/                     # React frontend
+├── docker/                     # MySQL / Redis Docker images and configs
+├── tutorials/                  # Colab notebooks (prediction & XAI visualization)
+├── docs/                       # README assets (architectures, ablation, benchmarks, samples, XAI results)
+├── .github/                    # CI workflows and repository settings
+├── docker-compose.yml          # Full stack (MySQL, Redis, FastAPI, Celery, React)
+├── train_eff_vit.py            # Training entrypoint for MS-EffViT
+├── train_eff_gcvit.py          # Training entrypoint for MS-EffGCViT
+├── pyproject.toml              # Package build config
+├── requirements.txt            # Dev environment dependencies
+├── README.md, README_KR.md, README_JP.md
+└── LICENSE
+```
+
 ## 🐳 Docker Quick Start
 
 Run the full stack (MySQL + Redis + FastAPI + Celery + React) with Docker Compose — no local Python/Node setup required.
@@ -223,10 +247,19 @@ We utilizes two distinct types of self-attention to capture both long-range and
 
 ## 🧬 Model Zoo
 
-| Model | Resolution | # Total Params(M) | # Backbone(M) | # L-ViT(M) | # H-ViT(M)  | FLOPs (G) | Model Config |
-| ----- | ---------- | -------------- | ----------- |------------- | ------------- | --------------  | ------- | 
-| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.7 | 3.6(41.4%) | 1.7(19.5%) | 3.3(37.9%) | 0.87 |  [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
-| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.3 | 27.3(54.3%) | 6.6(13.1%) | 16.1(32.0%) | 13.64 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| Model | Resolution | # Params(M) | # FLOPS(G) | Model Config |
+| ----- | ---------- | -------------- | ----------- | ------- | 
+| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.70 | 0.85 |  [spec](deepguard/config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
+| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.34 | 13.43 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+
+## 🕐 Model Inference Latency
+
+Latency is measured with batch size 1 on a single Tesla T4 and video (num_frames=20).
+
+| Model | Mode | Target Deployment | Latency (s) | Ratio |
+| ----- | ---- | ----------------- | ----------- | ----- |
+| ⚡ ms_eff_gcvit_b0 | Fast Mode | Edge / on-device deployment | 0.051 | 1.00× |
+| 🔥 ms_eff_gcvit_b5 | Pro Mode | Enterprise deployment | 0.479 | 9.41× |
 
 ## 🚀 Training
 
@@ -240,13 +273,6 @@ pip install -r requirements.txt
 
 We provide training scripts for both `ms_eff_vit` and `ms_eff_gcvit`. We recommend using **Google Colab** for free GPU access and **Weightes & Biases(W&B)** for experiment tracking
 
-#### 📊 Weight & Biases Experiments
-
-* **ms_eff_vit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_kodf)
-* **ms_eff_vit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_kodf)
-* **ms_eff_gcvit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_kodf)
-* **ms_eff_gcvit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_kodf)
-
 ```python
 !python -m train_eff_vit \ # train_eff_gcvit
     --root-dir DATA_ROOT \ 
@@ -601,6 +627,9 @@ The jupyter notebooks themselves can be found under the tutorials folder in the
 | 8 | [`Pytorch-Grad-Cam`](https://github.com/jacobgil/pytorch-grad-cam) | Advanced Visual Explanations for PyTorch Models |
 | 9 | [`DeepfakeBench`](https://github.com/SCLBD/DeepfakeBench) | A Comprehensive Benchmark of Deepfake Detection by SCLBD |
 | 10 | [`Effort-AIGI-Detection`](https://github.com/YZY-stack/Effort-AIGI-Detection) | Official Implementation of Effort (ICML 2025 Oral) by YZY-stack |
+| 11 | [`SIA`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/sia_detector.py) | An Information Theoretic Approach for Attention-Driven Face Forgery Detection (ECCV 2022), implemented in DeepfakeBench |
+| 12 | [`CORE`](https://github.com/niyunsheng/CORE) | Official Implementation of CORE: COnsistent REpresentation Learning for Face Forgery Detection (CVPRW 2022) by niyunsheng |
+| 13 | [`IID`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/iid_detector.py) | Implicit Identity Driven Deepfake Face Swapping Detection (CVPR 2023), implemented in DeepfakeBench |
 
 ## ⚖️ License 
 
diff --git a/README_JP.md b/README_JP.md
index ef7dea9a..f0a3a17e 100644
--- a/README_JP.md
+++ b/README_JP.md
@@ -25,7 +25,6 @@
   <img src="https://img.shields.io/badge/Model-MS--Eff--GCViT--B0%20%2F%20B5-orange?style=flat-square" alt="Models">
   <img src="https://img.shields.io/badge/Python-3.10+-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python">
   <img src="https://img.shields.io/badge/PyTorch-EE4C2C?style=flat-square&logo=pytorch&logoColor=white" alt="PyTorch">
-  <img src="https://img.shields.io/badge/W%26B-Recording-FFBE00?style=flat-square&logo=weightsandbiases&logoColor=white" alt="W&B">
 </p>
 
 <p align="center">
@@ -44,12 +43,14 @@
 
 ## 📌 目次
 
+- [🗂 プロジェクト構成](#-プロジェクト構成) - ディレクトリ構成の概要 (パッケージ、Webアプリ、ドキュメント)
 - [🐳 Docker Quick Start](#-docker-quick-start) - Docker Compose でフルスタック(MySQL, Redis, FastAPI, Celery, React)を起動
 - [🤗 今すぐ体験: Hugging Face Spaces](#-今すぐ体験-hugging-face-spaces) - インストール不要、ブラウザだけで試せる画像/動画/XAI ディープフェイク検出デモ
 - [📚 ディープフェイク動画ベンチマークデータセット](#-ディープフェイク動画ベンチマークデータセット) — 学習に使用した Celeb-DF-v2, FF++, KoDF データセットの概要。
 - [⚙️ データ準備](#データ準備) — YOLOv8 を用いた効率的な顔検出およびランドマーク抽出パイプライン。
 - [🏗 モデルアーキテクチャ](#-モデルアーキテクチャ) — ハイブリッド CNN-ViT (MS-EffViT & MS-EffGCViT) 設計の詳細。
 - [🧬 Model Zoo](#-model-zoo) — モデルバリアントごとのパラメータ数および演算量(FLOPs)の比較。
+- [🕐 Model Inference Latency](#-model-inference-latency) — Tesla T4 1基における Fast(b0)と Pro(b5)の推論速度の比較。
 - [🚀 学習](#-学習) - Google Colab および W&B を活用した段階的な学習スクリプト。
 - [📈 モデル評価](#-モデル評価) - ベンチマーク結果。
 - [💻 モデルの使い方](#-モデルの使い方) - `pip install deepguard` または Hugging Face Hub から学習済みモデルをロード
@@ -62,6 +63,29 @@
 
 ---
 
+## 🗂 プロジェクト構成
+
+```
+DeepGuard/
+├── deepguard/                  # コアパッケージ (モデル、レイヤー、データパイプライン、オプティマイザ、スケジューラ、ユーティリティ)
+├── preprocess/                 # 顔検出、クロップ、ランドマーク抽出パイプライン
+├── inference/                  # 画像 / 動画の予測
+├── explainability/             # デュアルブランチ CAM ベース XAI
+├── App/                        # FastAPI バックエンド (PyPI パッケージから除外)
+├── client/                     # React フロントエンド
+├── docker/                     # MySQL / Redis Docker イメージと設定
+├── tutorials/                  # Colab ノートブック (予測と XAI 可視化)
+├── docs/                       # README 用アセット (アーキテクチャ、ablation、ベンチマーク、サンプル、XAI 結果)
+├── .github/                    # CI ワークフローとリポジトリ設定
+├── docker-compose.yml          # フルスタック (MySQL, Redis, FastAPI, Celery, React)
+├── train_eff_vit.py            # MS-EffViT 学習エントリポイント
+├── train_eff_gcvit.py          # MS-EffGCViT 学習エントリポイント
+├── pyproject.toml              # パッケージビルド設定
+├── requirements.txt            # 開発環境の依存関係
+├── README.md, README_KR.md, README_JP.md
+└── LICENSE
+```
+
 ## 🐳 Docker Quick Start
 
 Docker Compose でフルスタック(MySQL + Redis + FastAPI + Celery + React)を起動します — ローカルに Python/Node 環境を用意する必要はありません。
@@ -221,10 +245,19 @@ DATA_ROOT/
 
 ## 🧬 Model Zoo
 
-| モデル名 | 解像度 | 総パラメータ(M) | バックボーン(M) | L-ViT(M) | H-ViT(M) | 演算量(FLOPs, G) | 設定ファイル |
-| ----- | ---------- | -------------- | ----------- |------------- | ------------- | --------------  | ------- | 
-| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.7 | 3.6(41.4%) | 1.7(19.5%) | 3.3(37.9%) | 0.87 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
-| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.3 | 27.3(54.3%) | 6.6(13.1%) | 16.1(32.0%) | 13.64 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| モデル名 | 解像度 | パラメータ(M) | 演算量(GFLOPS) | 設定ファイル |
+| ----- | ---------- | -------------- | ----------- | ------- | 
+| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.7 | 0.85 | [spec](deepguard/config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
+| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.3 | 13.43 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+
+## 🕐 Model Inference Latency
+
+推論速度は、Tesla T4 1基でバッチサイズ 1、動画(num_frames=20)を対象に測定しています。
+
+| モデル名 | モード | 展開対象 | 推論速度(s) | 比率 |
+| ----- | ---- | -------- | ----------- | ----- |
+| ⚡ ms_eff_gcvit_b0 | Fast Mode | エッジ / オンデバイス展開向け | 0.051 | 1.00× |
+| 🔥 ms_eff_gcvit_b5 | Pro Mode | エンタープライズ展開向け | 0.479 | 9.41× |
 
 ## 🚀 学習
 
@@ -238,13 +271,6 @@ pip install -r requirements.txt
 
 `ms_eff_vit` と `ms_eff_gcvit` の両方について学習スクリプトを提供します。無料の GPU 環境として **Google Colab** を、実験の記録およびトラッキングとして **Weights & Biases(W&B)** の使用を推奨します。
 
-#### 📊 Weight & Biases 実験結果
-
-* **ms_eff_vit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_kodf)
-* **ms_eff_vit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_kodf)
-* **ms_eff_gcvit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_kodf)
-* **ms_eff_gcvit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_kodf)
-
 ```python
 !python -m train_eff_vit \ # または train_eff_gcvit
     --root-dir DATA_ROOT \ 
@@ -593,6 +619,9 @@ Jupyterノートブックは、gitリポジトリのtutorialsフォルダ内に
 | 8 | [`Pytorch-Grad-Cam`](https://github.com/jacobgil/pytorch-grad-cam) | PyTorch モデルのための高度な視覚的説明ツール |
 | 9 | [`DeepfakeBench`](https://github.com/SCLBD/DeepfakeBench) | SCLBD による包括的なディープフェイク検出ベンチマーク |
 | 10 | [`Effort-AIGI-Detection`](https://github.com/YZY-stack/Effort-AIGI-Detection) | YZY-stack による Effort（ICML 2025 Oral）の公式実装 |
+| 11 | [`SIA`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/sia_detector.py) | An Information Theoretic Approach for Attention-Driven Face Forgery Detection（ECCV 2022）、DeepfakeBench 内の実装 |
+| 12 | [`CORE`](https://github.com/niyunsheng/CORE) | niyunsheng による CORE: COnsistent REpresentation Learning for Face Forgery Detection（CVPRW 2022）の公式実装 |
+| 13 | [`IID`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/iid_detector.py) | Implicit Identity Driven Deepfake Face Swapping Detection（CVPR 2023）、DeepfakeBench 内の実装 |
 
 ## ⚖️ ライセンス 
 
diff --git a/README_KR.md b/README_KR.md
index decec284..ab3c44a9 100644
--- a/README_KR.md
+++ b/README_KR.md
@@ -25,7 +25,6 @@
   <img src="https://img.shields.io/badge/Model-MS--Eff--GCViT--B0%20%2F%20B5-orange?style=flat-square" alt="Models">
   <img src="https://img.shields.io/badge/Python-3.10+-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python">
   <img src="https://img.shields.io/badge/PyTorch-EE4C2C?style=flat-square&logo=pytorch&logoColor=white" alt="PyTorch">
-  <img src="https://img.shields.io/badge/W%26B-Recording-FFBE00?style=flat-square&logo=weightsandbiases&logoColor=white" alt="W&B">
 </p>
 
 <p align="center">
@@ -44,12 +43,14 @@
 
 ## 📌 목차
 
+- [🗂 프로젝트 구조](#-프로젝트-구조) - 디렉터리 구조 개요 (패키지, 웹 앱, 문서)
 - [🐳 Docker Quick Start](#-docker-quick-start) - Docker Compose로 전체 스택(MySQL, Redis, FastAPI, Celery, React) 실행
 - [🤗 지금 바로 체험하기: Hugging Face Spaces](#-지금-바로-체험하기-hugging-face-spaces) - 설치 없이 브라우저에서 바로 체험하는 이미지/비디오/XAI 딥페이크 탐지 데모
 - [📚 딥페이크 비디오 벤치마크 데이터셋](#-딥페이크-비디오-벤치마크-데이터셋) — 학습에 사용된 Celeb-DF-v2, FF++, KoDF 데이터셋 개요.
 - [⚙️ 데이터 준비](#데이터-준비) — YOLOv8을 이용한 효율적인 얼굴 검출 및 랜드마크 추출 파이프라인.
 - [🏗 모델 구조](#-모델-구조) — 하이브리드 CNN-ViT (MS-EffViT & MS-EffGCViT) 설계 상세.
 - [🧬 모델 주(Model Zoo)](#-model-zoo) — 모델 변체별 파라미터 수 및 연산량(FLOPs) 비교.
+- [🕐 Model Inference Latency](#-model-inference-latency) — Tesla T4 1대 기준 Fast(b0) vs Pro(b5) 추론 속도 비교.
 - [🚀 학습](#-학습) - Google Colab 및 W&B를 활용한 단계별 학습 스크립트.
 - [📈 모델 평가](#-모델-평가) - 벤치마크 결과.
 - [💻 모델 사용법](#-모델-사용법) - `pip install deepguard` 또는 Hugging Face Hub를 통한 사전 학습 모델 로드
@@ -62,6 +63,29 @@
 
 ---
 
+## 🗂 프로젝트 구조
+
+```
+DeepGuard/
+├── deepguard/                  # 핵심 패키지 (모델, 레이어, 데이터 파이프라인, 옵티마이저, 스케줄러, 유틸)
+├── preprocess/                 # 얼굴 검출, 크롭 및 랜드마크 추출 파이프라인
+├── inference/                  # 이미지 / 비디오 예측
+├── explainability/             # 듀얼 브랜치 CAM 기반 XAI
+├── App/                        # FastAPI 백엔드 (PyPI 패키지에서 제외)
+├── client/                     # React 프론트엔드
+├── docker/                     # MySQL / Redis Docker 이미지 및 설정
+├── tutorials/                  # Colab 노트북 (예측 및 XAI 시각화)
+├── docs/                       # README 자료 (아키텍처, ablation, 벤치마크, 샘플, XAI 결과)
+├── .github/                    # CI 워크플로 및 저장소 설정
+├── docker-compose.yml          # 전체 스택 (MySQL, Redis, FastAPI, Celery, React)
+├── train_eff_vit.py            # MS-EffViT 학습 엔트리포인트
+├── train_eff_gcvit.py          # MS-EffGCViT 학습 엔트리포인트
+├── pyproject.toml              # 패키지 빌드 설정
+├── requirements.txt            # 개발 환경 의존성
+├── README.md, README_KR.md, README_JP.md
+└── LICENSE
+```
+
 ## 🐳 Docker Quick Start
 
 Docker Compose로 전체 스택(MySQL + Redis + FastAPI + Celery + React)을 실행합니다 — 로컬에 Python/Node 환경을 따로 설치할 필요가 없습니다.
@@ -221,10 +245,19 @@ DATA_ROOT/
 
 ## 🧬 Model Zoo
 
-| 모델명 | 해상도 | 총 파라미터(M) | 백본(M) | L-ViT(M) | H-ViT(M) | 연산량(FLOPs, G) | 설정 파일 |
-| ----- | ---------- | -------------- | ----------- |------------- | ------------- | --------------  | ------- | 
-| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.7 | 3.6(41.4%) | 1.7(19.5%) | 3.3(37.9%) | 0.87 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
-| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.3 | 27.3(54.3%) | 6.6(13.1%) | 16.1(32.0%) | 13.64 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| 모델명 | 해상도 | 파라미터(M) | 연산량(GFLOPS) | 설정 파일 |
+| ----- | ---------- | -------------- | ----------- | ------- | 
+| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.70 | 0.85 | [spec](deepguard/config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
+| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.34 | 13.43 | [spec](deepguard/config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+
+## 🕐 Model Inference Latency
+
+추론 속도는 Tesla T4 GPU 1대에서 배치 크기 1, 비디오(num_frames=20) 기준으로 측정했습니다.
+
+| 모델명 | 모드 | 배포 대상 | 추론 속도(s) | 비율 |
+| ----- | ---- | -------- | ----------- | ----- |
+| ⚡ ms_eff_gcvit_b0 | Fast Mode | 엣지 / 온디바이스 배포용 | 0.051 | 1.00× |
+| 🔥 ms_eff_gcvit_b5 | Pro Mode | 엔터프라이즈 배포용 | 0.479 | 9.41× |
 
 ## 🚀 학습
 
@@ -238,13 +271,6 @@ pip install -r requirements.txt
 
 `ms_eff_vit` 및 `ms_eff_gcvit` 모두에 대한 학습 스크립트를 제공합니다. 무료 GPU 환경을 위해 **Google Colab**을, 실험 기록 및 트래킹을 위해 **Weights & Biases(W&B)** 사용을 권장합니다.
 
-#### 📊 Weight & Biases 실험 결과
-
-* **ms_eff_vit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b0_kodf)
-* **ms_eff_vit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_vit_b5_kodf)
-* **ms_eff_gcvit_b0:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b0_kodf)
-* **ms_eff_gcvit_b5:** [Celeb-DF-v2 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_celeb_df_v2) | [FaceForensics++ 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_ff++) | [KoDF 🚀](https://wandb.ai/origin6165/ms_eff_gcvit_b5_kodf)
-
 ```python
 !python -m train_eff_vit \ # 또는 train_eff_gcvit
     --root-dir DATA_ROOT \ 
@@ -593,6 +619,9 @@ result = explainer.display_heatmap_bbox_on_image(
 | 8 | [`Pytorch-Grad-Cam`](https://github.com/jacobgil/pytorch-grad-cam) | PyTorch 모델을 위한 고급 시각적 설명 도구 |
 | 9 | [`DeepfakeBench`](https://github.com/SCLBD/DeepfakeBench) | SCLBD의 포괄적인 딥페이크 탐지 벤치마크 |
 | 10 | [`Effort-AIGI-Detection`](https://github.com/YZY-stack/Effort-AIGI-Detection) | YZY-stack의 Effort(ICML 2025 Oral) 공식 구현체 |
+| 11 | [`SIA`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/sia_detector.py) | An Information Theoretic Approach for Attention-Driven Face Forgery Detection(ECCV 2022), DeepfakeBench 내 구현 |
+| 12 | [`CORE`](https://github.com/niyunsheng/CORE) | niyunsheng의 CORE: COnsistent REpresentation Learning for Face Forgery Detection(CVPRW 2022) 공식 구현체 |
+| 13 | [`IID`](https://github.com/SCLBD/DeepfakeBench/blob/main/training/detectors/iid_detector.py) | Implicit Identity Driven Deepfake Face Swapping Detection(CVPR 2023), DeepfakeBench 내 구현 |
 
 ## ⚖️ 라이선스 
 
diff --git a/deepguard/MS_EffGCViT.md b/deepguard/MS_EffGCViT.md
index 169740e9..3c064367 100644
--- a/deepguard/MS_EffGCViT.md
+++ b/deepguard/MS_EffGCViT.md
@@ -15,7 +15,7 @@ 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**
 
 
-<img src="../docs/benchmarks/celeb_df_v2_gcvit.png" width="900">
+<img src="../docs/benchmarks/gcvit_params_acc.png" width="900">
 
 ## 💥 News 💥
 - [**09.05.2026**] 🔥 Released **HIGH-Eff-GCVIT-B5** (384X384) weights for **Celeb DF(V2)**, **FaceForensics++**, **KoDF**
@@ -32,36 +32,42 @@ This model is a **frame-level** and **spatial-domain** architecture, designed to
 
 **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.
+deployment and **Pro (b5)** for enterprise-grade accuracy.
 
-<p align="center">
-  <img src="../docs/benchmarks/gcvit_summary_bars.png" width="100%">
-</p>
+**📊 Model Performance Results on Celeb-DF (v2) Dataset**
 
-> 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.
+| Model | Test@Acc | Test@Auc | #Params(M) | Params Ratio | #FLOPS(G) | FLOPS Ratio |
+| --- | --- | --- | --- | --- | --- | --- |
+| **ms_eff_gcvit_b0** | 98.4% | 99.7% | 8.7 | 1x | 0.85 | 1x |
+| sia (eccv,2022) | 99.0% | 99.8% | 17.57 | 2x | 2.05 | 2.4x |
+| core (cvpr,2022) | 99.6% | 99.7% | 20.81 | 2.4x | 8.40 | 9.9x |
+| **ms_eff_gcvit_b5** | **99.8%** | 99.8% | 50.34 | 5.8x | 13.43 | 15.8x |
+| iid (cvpr,2023) | **99.8%** | **99.9%** | 65.5 | 7.5x | 12.7 | 14.9x |
+| effort (icml, 2025) | **99.8%** | 99.7% | 303.4 | 34.9x | 155.9 | 183.4x |
 
-<details>
-<summary><b>📊 Celeb-DF (v2) — Accuracy & Efficiency</b></summary>
-<br>
-<img src="../docs/benchmarks/celeb_df_v2_gcvit_2.png" width="100%">
 
-</details>
+**📊 Model Performance Results on FaceForensics++ Dataset**
 
-<details>
-<summary><b>📊 FaceForensics++ — Accuracy & Efficiency</b></summary>
-<br>
-<img src="../docs/benchmarks/ff_gcvit.png" width="100%">
+| Model | Test@Acc | Test@Auc | #Params(M) | Params Ratio | #FLOPS(G) | FLOPS Ratio |
+| --- | --- | --- | --- | --- | --- | --- |
+| **ms_eff_gcvit_b0** | 98.1% | 99.7% | 8.7 | 1x | 0.85 | 1x |
+| sia (ECCV,2022) | 95.0% | 97.5% | 17.57 | 2x | 2.05 | 2.4x |
+| core (CVPR,2022) | 98.3% | **99.8%** | 20.81 | 2.4x | 8.40 | 9.9x |
+| **ms_eff_gcvit_b5** | **98.5%** | 99.7% | 50.34 | 5.8x | 13.43 | 15.8x |
+| iid (CVPR,2023) | 97.3% | 99.6% | 65.5 | 7.5x | 12.7 | 14.9x |
+| effort (ICML, 2025) | 98.4% | **99.8%** | 303.4 | 34.9x | 155.9 | 183.4x |
 
-</details>
+**📊 Model Performance Results on KoDF Dataset**
 
-<details>
-<summary><b>📊 KoDF Competition — Accuracy Ranking</b></summary>
-<br>
-<img src="../docs/benchmarks/kodf_gcvit.png" width="100%">
+| Model | Test@Acc | Test@Auc | #Params(M) | Params Ratio | #FLOPS(G) | FLOPS Ratio |
+| --- | --- | --- | --- | --- | --- | --- |
+| **ms_eff_gcvit_b0** | 96.6% | 97.9% | 8.7 | 1x | 0.85 | 1x |
+| sia (ECCV,2022) | 97.0% | 98.8% | 17.57 | 2x | 2.05 | 2.4x |
+| core (CVPR,2022) | 96.3% | 99.2% | 20.81 | 2.4x | 8.40 | 9.9x |
+| **ms_eff_gcvit_b5** | 97.9% | 98.3% | 50.34 | 5.8x | 13.43 | 15.8x |
+| iid (CVPR,2023) | 96.5% | 97.6% | 65.5 | 7.5x | 12.7 | 14.9x |
+| effort (ICML, 2025) | **99.0%** | **99.5%** | 303.4 | 34.9x | 155.9 | 183.4x |
 
-</details>
 
 ## Model Indroduction
 
@@ -109,28 +115,20 @@ Modern DeepFakes can leave very localized forgery region. To Capture this, we ad
 
 ## 📊 Model Zoo
 
-| Model | Resolution | # Total Params(M) | # Backbone(M) | # L-ViT(M) | # H-ViT(M)  | FLOPs (G) | Model Config |
-| ----- | ---------- | -------------- | ----------- |------------- | ------------- | --------------  | ------- | 
-| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.7 | 3.6(41.4%) | 1.7(19.5%) | 3.3(37.9%) | 0.87 |  [spec](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
-| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.3 | 27.3(54.3%) | 6.6(13.1%) | 16.1(32.0%) | 13.64 | [spec](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| Model | Resolution | # Params(M) | # FLOPS(G) | Model Config |
+| ----- | ---------- | -------------- | ----------- | ------- | 
+| ⚡ ms_eff_gcvit_b0 | 224 X 224 | 8.70 | 0.85 | [spec](./config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
+| 🔥 ms_eff_gcvit_b5 | 384 X 384 | 50.34 | 13.43 | [spec](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
 
 
-## 🛠 Model Variants
+## 🕐 Model Inference Latency
 
-**⚡ ms_eff_gcvit_b0 (Fast Mode / Mobile)**: Efficiency at the Edge
-- Optimized for **real-time inference** and mobile deployment.
+Latency is measured with batch size 1 on a single Tesla T4 and video (num_frames=20).
 
-**🔥 ms_eff_gcvit_b5 (Pro Mode / Enterprise)**: Uncompromising Precision
-- Engineered for high-fidelity analysis and enterprise-grade accuracy.
-
-## ⚙️ Model Weight Initialization
-The model incorporates a hybrid initialization strategy to leverage pre-trained features while ensuring stable convergence of the transformer components
-
-> **Backbone**: **ImageNet-1K Pretraiend Weights**(EfficientNet)  
-
-> **L-GCViT / H-GCViT / Head**: Truncated Normal( `std=0.02` )
- 
-> **No Weight Decay**: `relative_position_bias_table`, `bn`, `norm`
+| Model | Mode | Target Deployment | Latency (s) | Ratio |
+| ----- | ---- | ----------------- | ----------- | ----- |
+| ⚡ ms_eff_gcvit_b0 | Fast Mode | Edge / on-device deployment | 0.051 | 1.00× |
+| 🔥 ms_eff_gcvit_b5 | Pro Mode | Enterprise deployment | 0.479 | 9.41× |
 
 
 ## DeepFake Video Benchmarks
@@ -143,24 +141,24 @@ The model incorporates a hybrid initialization strategy to leverage pre-trained
 
 **Celeb DF(v2) Pretrained Models**
 
-| Model Variant | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| Model | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
 | ------------- | -------- | -------- | ---------- | -------- | ----- |
 | ms_eff_gcvit_b0 | 0.9842 | 0.9965 | 0.0283 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b0_celeb_df_v2.bin) | [recipe](./config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
 | ms_eff_gcvit_b5 | 0.9981 | 0.9984 | 0.0089 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b5_celeb_df_v2.bin) | [recipe](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
 
 **FaceForensics++ Pretrained Models**
 
-| Model Variant | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| Model | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
 | ------------- | -------- | -------- | ---------- | -------- | ------ |
-| ms_eff_gcvit_b0 | 0.9808 | 0.9969 | 0.0637| [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b0_ff++.bin) | [recipe](./config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
-| ms_eff_gcvit_b5 | 0.9850 | 0.9974 | 0.0492 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b5_ff++.bin) | [recipe](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| ms_eff_gcvit_b0 | 0.9808 | 0.9969 | 0.0637| [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b0_ff++.bin) | [recipe](./config/ms_eff_gcvit_b0/ff++.yaml) |
+| ms_eff_gcvit_b5 | 0.9850 | 0.9974 | 0.0492 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.1.0/ms_eff_gcvit_b5_ff++.bin) | [recipe](./config/ms_eff_gcvit_b5/ff++.yaml) |
 
 **KoDF Pretrained Models**
 
-| Model Variant | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| Model | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
 | ------------- | -------- | -------- | ---------- | -------- | ------ |
-| ms_eff_gcvit_b0 | 0.9655 | 0.9792 | 0.1237 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.2.0/ms_eff_gcvit_b0_kodf.bin) | [recipe](./config/ms_eff_gcvit_b0/celeb_df_v2.yaml) |
-| ms_eff_gcvit_b5 | 0.9850 | 0.9974 | 0.0492 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.2.0/ms_eff_gcvit_b5_kodf.bin) | [recipe](./config/ms_eff_gcvit_b5/celeb_df_v2.yaml) |
+| ms_eff_gcvit_b0 | 0.9655 | 0.9792 | 0.1237 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.2.0/ms_eff_gcvit_b0_kodf.bin) | [recipe](./config/ms_eff_gcvit_b0/kodf.yaml) |
+| ms_eff_gcvit_b5 | 0.9792 | 0.9831 | 0.0692 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v0.2.0/ms_eff_gcvit_b5_kodf.bin) | [recipe](./config/ms_eff_gcvit_b5/kodf.yaml) |
 
 ## Usage
 
diff --git a/docs/ablation/ablation_celeb_df_v2.webp b/docs/ablation/ablation_celeb_df_v2.webp
new file mode 100644
index 00000000..2dde6eca
Binary files /dev/null and b/docs/ablation/ablation_celeb_df_v2.webp differ
diff --git a/docs/ablation/ablation_feature_map.png b/docs/ablation/ablation_feature_map.png
new file mode 100644
index 00000000..53424f32
Binary files /dev/null and b/docs/ablation/ablation_feature_map.png differ
diff --git a/docs/ablation/ablation_ff++.webp b/docs/ablation/ablation_ff++.webp
new file mode 100644
index 00000000..081ec75f
Binary files /dev/null and b/docs/ablation/ablation_ff++.webp differ
diff --git a/docs/ablation/ablation_kodf.webp b/docs/ablation/ablation_kodf.webp
new file mode 100644
index 00000000..26bf888e
Binary files /dev/null and b/docs/ablation/ablation_kodf.webp differ
diff --git a/docs/benchmarks/celeb_df_v2_gcvit.png b/docs/benchmarks/celeb_df_v2_gcvit.png
deleted file mode 100644
index 77190479..00000000
Binary files a/docs/benchmarks/celeb_df_v2_gcvit.png and /dev/null differ
diff --git a/docs/benchmarks/celeb_df_v2_gcvit_2.png b/docs/benchmarks/celeb_df_v2_gcvit_2.png
deleted file mode 100644
index 5cc437f2..00000000
Binary files a/docs/benchmarks/celeb_df_v2_gcvit_2.png and /dev/null differ
diff --git a/docs/benchmarks/ff_gcvit.png b/docs/benchmarks/ff_gcvit.png
deleted file mode 100644
index 9a70a73b..00000000
Binary files a/docs/benchmarks/ff_gcvit.png and /dev/null differ
diff --git a/docs/benchmarks/gcvit_params_acc.png b/docs/benchmarks/gcvit_params_acc.png
new file mode 100644
index 00000000..a1095322
Binary files /dev/null and b/docs/benchmarks/gcvit_params_acc.png differ
diff --git a/docs/benchmarks/gcvit_summary_bars.png b/docs/benchmarks/gcvit_summary_bars.png
deleted file mode 100644
index ee8d30b4..00000000
Binary files a/docs/benchmarks/gcvit_summary_bars.png and /dev/null differ
diff --git a/docs/benchmarks/kodf_gcvit.png b/docs/benchmarks/kodf_gcvit.png
deleted file mode 100644
index d163b389..00000000
Binary files a/docs/benchmarks/kodf_gcvit.png and /dev/null differ
diff --git a/explainability/explainer/base_explainer.py b/explainability/explainer/base_explainer.py
index bb815106..da4e93b4 100644
--- a/explainability/explainer/base_explainer.py
+++ b/explainability/explainer/base_explainer.py
@@ -11,11 +11,17 @@
 class BaseExplainer:
     def __init__(
         self,
-        model_name: Literal["ms_eff_vit_b0", "ms_eff_vit_b5", "ms_eff_gcvit_b0", "ms_eff_gcvit_b5"],
+        model_name: Literal[
+            "ms_eff_vit_b0", "ms_eff_vit_b5", "ms_eff_gcvit_b0", "ms_eff_gcvit_b5",
+            # ablation models
+            "eff_b0", "eff_b5",
+            "low_eff_gcvit_b0", "low_eff_gcvit_b5",
+            "high_eff_gcvit_b0", "high_eff_gcvit_b5",
+        ],
         dataset: Literal["celeb_df_v2", "ff++", "kodf"], 
         margin_ratio: float = 0.2,
         conf_thres: float = 0.5,
-        branch_level: Literal["low", "high"] = "high",
+        branch_level: Literal["low", "high", "both"] = "high", # "both"는 ms_eff_* 모델만 가능
         l_stage_idx: int = -1, # gcvit low branch stage index (0~3)
         block_idx: int = -1, # vit, gcvit block index
         model: Optional[torch.nn.Module] = None,
@@ -32,8 +38,15 @@ def __init__(
                 
         # Parse Model Metadata
         self.model_variant = model_name.split("_")[-1] # b0, b5
-        self.transformer_type = model_name.split("_")[-2] # vit, gcvit 
-                
+        # vit, gcvit, cnn (eff_b0/eff_b5: transformer 없는 EfficientNet baseline)
+        if "gcvit" in model_name:
+            self.transformer_type = "gcvit"
+        elif "vit" in model_name:
+            self.transformer_type = "vit"
+        else:
+            self.transformer_type = "cnn"
+        self._validate_branch_level()
+
         # Model & Tools setup
         self.model = model if model is not None else timm.create_model(
         model_name, pretrained=True, dataset=dataset
@@ -54,8 +67,34 @@ def _get_reshape_fn(self) -> Callable[[torch.Tensor], torch.Tensor]:
         Returns:
             Callable: 3D/4D 텐서를 GradCAM 연산에 적합한 형태로 변환하는 함수.
         """
-        return reshape_transform_vit if self.transformer_type == "vit" else reshape_transform_gcvit    
-    
+        if self.transformer_type == "cnn":
+            return None  # backbone feature map은 이미 (B,C,H,W)
+        return reshape_transform_vit if self.transformer_type == "vit" else reshape_transform_gcvit
+
+    def _validate_branch_level(self) -> None:
+        """Ablation 모델이 실제로 가진 branch와 branch_level이 일치하는지 검증한다.
+
+        Raises:
+            ValueError: low_/high_ 단일 branch 모델에 존재하지 않는 branch를 요청한 경우 발생.
+        """
+        if self.transformer_type == "cnn" and self.branch_level != "high":
+            raise ValueError(f"{self.model_name}은 마지막 feature map만 사용하므로 branch_level='high'만 가능합니다.")
+        if self.model_name.startswith("low_") and self.branch_level != "low":
+            raise ValueError(f"{self.model_name}은 low-level branch만 가지므로 branch_level='low'만 가능합니다.")
+        if self.model_name.startswith("high_") and self.branch_level != "high":
+            raise ValueError(f"{self.model_name}은 high-level branch만 가지므로 branch_level='high'만 가능합니다.")
+
+    def _get_backbone_layer(self, feature_idx: int) -> torch.nn.Module:
+        """EfficientNet backbone에서 feature_idx 단계의 출력 모듈을 찾는다."""
+        backbone = self.model.feat_extractor.backbone
+        name = backbone.feature_info[feature_idx]["module"]  # e.g. "blocks.6"
+        modules = dict(backbone.named_modules())
+        # timm FeatureListNet은 이름의 '.'를 '_'로 바꾸는 경우가 있어 둘 다 시도
+        for key in (name, name.replace(".", "_")):
+            if key in modules:
+                return modules[key]
+        raise ValueError(f"backbone에서 '{name}' 모듈을 찾을 수 없습니다.")
+
     def _resolve_target_layers(self) -> list[torch.nn.Module]:
         """시각화(XAI)의 대상이 될 Target Layer들을 선택한다.
 
@@ -67,14 +106,18 @@ def _resolve_target_layers(self) -> list[torch.nn.Module]:
         """
         layers = []
 
-        if self.transformer_type == "gcvit":
-            n_stages  = len(self.model.l_gcvit.stages)
-            stage_idx = self.l_stage_idx if self.l_stage_idx >= 0 else n_stages + self.l_stage_idx
-
-            if not (0 <= stage_idx < n_stages):
-                raise ValueError(f"l_stage_idx={self.l_stage_idx} 범위 초과 (stage 수: {n_stages})")
+        if self.transformer_type == "cnn":
+            # eff_b0/eff_b5: transformer가 없으므로 backbone의 마지막 feature map을 직접 사용
+            layers.append(self._get_backbone_layer(self.model.h_meta["feature_idx"]))
 
+        elif self.transformer_type == "gcvit":
             if self.branch_level in ("low", "both"):
+                n_stages  = len(self.model.l_gcvit.stages)
+                stage_idx = self.l_stage_idx if self.l_stage_idx >= 0 else n_stages + self.l_stage_idx
+
+                if not (0 <= stage_idx < n_stages):
+                    raise ValueError(f"l_stage_idx={self.l_stage_idx} 범위 초과 (stage 수: {n_stages})")
+
                 n_blocks = len(self.model.l_gcvit.stages[stage_idx].blocks)
                 blk      = self._resolve_block_idx(self.block_idx, n_blocks, f"l_gcvit stage[{stage_idx}]")
                 layers.append(self.model.l_gcvit.stages[stage_idx].blocks[blk].norm2)
diff --git a/inference/Ablation.md b/inference/Ablation.md
new file mode 100644
index 00000000..b5788f39
--- /dev/null
+++ b/inference/Ablation.md
@@ -0,0 +1,78 @@
+# Ablation Study
+
+<p align="center">
+  <img src="../docs/architectures/dual_branch.gif" alt="Dual-Branch architecture" width="100%">
+</p>
+
+- **![](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
+
+- **![](https://img.shields.io/badge/High_level_Branch-red?style=flat-square) (_Global Features_)**: Low-Resolution feature maps are extracted from deeper blocks(`h_block_idx`) to analyze overall lighting, shadows, and structural consistency.
+
+- **Feature Fusion**: The Outputs from both branches (`L-GCViT and H-GCViT`) are fused to make a comprehensive decision based on both local and global context.
+
+## Ablation Model Zoo
+
+| Model | Branch | Resolution | # Params(M) | #FLOPs(G) |
+| :--- | :---: | :---: | :---: | :---: |
+| eff_b0 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 224 × 224 | 3.65M | 0.76G |
+| eff_b5 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 384 × 384 | 27.42M | 13.79G |
+| low_eff_gcvit_b0 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 224 × 224 | 5.30M | 0.79G |
+| low_eff_gcvit_b5 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 384 × 384 | 33.92M | 10.70G |
+| high_eff_gcvit_b0 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 224 × 224 | 6.94M | 1.08G |
+| high_eff_gcvit_b5 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 384 × 384 | 43.51M | 18.59G |
+
+## Ablation Feature map
+
+<p align="center">
+  <img src="../docs/ablation/ablation_feature_map.png" alt="Ablation Study: Feature Map" width="100%">
+</p>
+
+# Ablation Evaluation
+
+### Celeb DF(V2)
+
+<p align="center">
+  <img src="../docs/ablation/ablation_celeb_df_v2.webp" alt="Celeb-DF(V2) Branch-Ablation Pareto plots" width="100%">
+</p>
+
+| Model | Branch | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| ------------- | -------- | -------- | -------- | ---------- | -------- | ----- |
+| eff_b0 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9643 | 0.9754 | 0.0977 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b0_celeb_df_v2.bin) | [recipe](../deepguard/config/eff_b0/celeb_df_v2.yaml) |
+| eff_b5 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9768 | 0.9837 | 0.0676 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b5_celeb_df_v2.bin) | [recipe](../deepguard/config/eff_b5/celeb_df_v2.yaml) |
+| low_eff_gcvit_b0 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.9788 | 0.9837 | 0.0658 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b0_celeb_df_v2.bin) | [recipe](../deepguard/config/low_eff_gcvit_b0/celeb_df_v2.yaml) |
+| low_eff_gcvit_b5 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.9865 | 0.9987 | 0.0150 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b5_celeb_df_v2.bin) | [recipe](../deepguard/config/low_eff_gcvit_b5/celeb_df_v2.yaml) |
+| high_eff_gcvit_b0 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9807 | 0.9884 | 0.0343 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b0_celeb_df_v2.bin) | [recipe](../deepguard/config/high_eff_gcvit_b0/celeb_df_v2.yaml) |
+| high_eff_gcvit_b5 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9942 | 0.9977 | 0.0101 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b5_celeb_df_v2.bin) | [recipe](../deepguard/config/high_eff_gcvit_b5/celeb_df_v2.yaml) |
+
+
+### FaceForensics++
+
+<p align="center">
+  <img src="../docs/ablation/ablation_ff++.webp" alt="FaceForensics++ Branch-Ablation Pareto plots" width="100%">
+</p>
+
+| Model | Branch | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| ------------- | -------- | -------- | -------- | ---------- | -------- | ------ |
+| eff_b0 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9202 | 0.9497 | 0.1523 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b0_ff++.bin) | [recipe](../deepguard/config/eff_b0/ff++.yaml) |
+| eff_b5 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9660 | 0.9831 | 0.0842 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b5_ff++.bin) | [recipe](../deepguard/config/eff_b5/ff++.yaml) |
+| low_eff_gcvit_b0 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.8033 | 0.9257 | 0.4148 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b0_ff++.bin) | [recipe](../deepguard/config/low_eff_gcvit_b0/ff++.yaml) |
+| low_eff_gcvit_b5 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.8700 | 0.9486 | 0.3365 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b5_ff++.bin) | [recipe](../deepguard/config/low_eff_gcvit_b5/ff++.yaml) |
+| high_eff_gcvit_b0 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9783 | 0.9886 | 0.0767 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b0_ff++.bin) | [recipe](../deepguard/config/high_eff_gcvit_b0/ff++.yaml) |
+| high_eff_gcvit_b5 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9808 | 0.9956 | 0.0606 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b5_ff++.bin) | [recipe](../deepguard/config/high_eff_gcvit_b5/ff++.yaml) |
+
+
+### Kodf 
+
+<p align="center">
+  <img src="../docs/ablation/ablation_kodf.webp" alt="KoDF Branch-Ablation Pareto plots" width="100%">
+</p>
+
+| Model | Branch | Test@Acc | Test@Auc | Test@log_loss | Download | Train Config |
+| ------------- | -------- | -------- | -------- | ---------- | -------- | ------ |
+| eff_b0 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9180 | 0.9520 | 0.2010 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b0_kodf.bin) | [recipe](../deepguard/config/eff_b0/kodf.yaml) |
+| eff_b5 | ![](https://img.shields.io/badge/No_Branch-lightgrey?style=flat-square) | 0.9601 | 0.9770 | 0.1120 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/eff_b5_kodf.bin) | [recipe](../deepguard/config/eff_b5/kodf.yaml) |
+| low_eff_gcvit_b0 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.8684 | 0.9481 | 0.2843 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b0_kodf.bin) | [recipe](../deepguard/config/low_eff_gcvit_b0/kodf.yaml) |
+| low_eff_gcvit_b5 | ![](https://img.shields.io/badge/Low_Branch-blue?style=flat-square) | 0.8808 | 0.9596 | 0.2381 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/low_eff_gcvit_b5_kodf.bin) | [recipe](../deepguard/config/low_eff_gcvit_b5/kodf.yaml) |
+| high_eff_gcvit_b0 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9462 | 0.9696 | 0.1761 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b0_kodf.bin) | [recipe](../deepguard/config/high_eff_gcvit_b0/kodf.yaml) |
+| high_eff_gcvit_b5 | ![](https://img.shields.io/badge/High_Branch-red?style=flat-square) | 0.9745 | 0.9911 | 0.0844 | [model](https://github.com/HanMoonSub/DeepGuard/releases/download/v1.1.0/high_eff_gcvit_b5_kodf.bin) | [recipe](../deepguard/config/high_eff_gcvit_b5/kodf.yaml) |
+
diff --git a/inference/README.md b/inference/README.md
index 2ee59dd9..6edf569b 100644
--- a/inference/README.md
+++ b/inference/README.md
@@ -1,23 +1,29 @@
-# 🕵️‍♂️ Inference Pipeline & Strategy
-The inference pipeline of this project is designed for real-world deepfake detection scenarios. Our strategy focuses on **capturing subtle forgery artifacts while ensuring highly stable and reliable predictions.**
+# Inference Strategy
 
-## 🎯 Label
+## Label
 
 Label Mapping: **`Fake: 1`** | **`Real: 0`**
 
-In the deepfake detection domain, misclassifying a fake video as real (False Negative) poses a significantly greater threat than misclassifying a real video as fake (False Positive).
-
 ## Test Time Augmentation(TTA)
 
-Deepfake artifacts often consist of subtle, pixel-level manipulations. Applying heavy augmentation can inadvertently destroy these crucial forensic clues. So We applied a **`simple Horizontal Flip (p=0.5).`**
+Deepfake artifacts often consist of subtle, pixel-level manipulations. Applying heavy augmentation can inadvertently destroy these crucial forensic clues. So We applied a **`simple Horizontal Flip (p=0.5).`** TTA is controlled by `tta_hflip` in `get_test_transforms` (`deepguard/data/transforms.py`):
+
+```python
+from deepguard.data import get_test_transforms
+
+# Horizontal Flip TTA (p=0.5)
+transforms = get_test_transforms(img_size=[224, 224], tta_hflip=0.5)
+```
+
+Internally, `get_test_transforms` resizes the image, appends `A.HorizontalFlip(p=tta_hflip)` only when `tta_hflip != 0`, and then applies normalization.
 
 ## Number of Frames & Face Crop
 
-To ensure consistent and fair evaluation across diverse benchmarks (including Celeb-DF-v2, KoDF, and FaceForensics++), we standardized the frame extraction and face cropping parameters during inference:
+To ensure consistent and fair evaluation across diverse benchmarks (including **_Celeb-DF-v2_**, **_KoDF_**, and **_FaceForensics++_**), we standardized the frame extraction and face cropping parameters during inference:
 
-- **`Fixed Number of Frames`**: We uniformly sample exactly `20 frames` per video. This guarantees a balanced computational load while maintaining sufficient temporal coverage across all datasets.
+**`Fixed Number of Frames`**: We uniformly sample exactly **`20 frames`** per video. This guarantees a balanced computational load while maintaining sufficient temporal coverage across all datasets.
 
-- **`Cropping Margin Ratio`**: A `20% margin` (margin_ratio=0.2) is applied around the detected face bounding box. This optimal margin ensures that blending boundaries and facial contours—crucial areas for deepfake artifacts—are fully preserved without including excessive background noise.
+**`Cropping Margin Ratio`**: A **`20% margin`** (margin_ratio=0.2) is applied around the detected face bounding box. This optimal margin ensures that blending boundaries and facial contours—crucial areas for deepfake artifacts—are fully preserved without including excessive background noise.
 
 ## Frame Aggregation Strategy
 
@@ -55,38 +61,47 @@ def heuristic_aggregation(preds: np.ndarray) -> float:
     else:
         return float(np.mean(preds))
 ```
-## 📚 DeepFake Video Benchmark Datasets
 
-To evaluate the generalization and robustness of our deepfake detection model, we utilize three large-scale, widely recognized benchmark datasets. Each dataset presents unique challenges and covers different types of forgery methods.
 
-| Dataset | Real Videos | Fake Videos | Year | Participants | Description (Paper Title) | Details |
-| :--- | :---: | :---: | :---: | :---: | :--- | :---: |
-| **Celeb-DF-v2** | 890 | 5,639 | 2019 | 59 | *A Large-scale Challenging Dataset for DeepFake Forensics* | [🔗 Readme](../preprocess/celeb_df_v2/README.md) |
-| **FaceForensics++** | 1,000 | 6,000 | 2019 | 1,000 | *Learning to Detect Manipulated Facial Images* | [🔗 Readme](../preprocess/ff++/README.md) |
-| **KoDF** | 62,166 | 175,776 | 2020 | 400 | *Large-Scale Korean Deepfake Detection Dataset* | [🔗 Readme](../preprocess/kodf/README.md) |
+# In-domain Evaluation
 
-> **Note:** For detailed information on data preprocessing, frame extraction, and formatting for each dataset, please refer to the respective `Readme` links in the table above.
+Each model is trained and evaluated on the same dataset: it is tested on the held-out test split of the dataset it was trained on. 
 
-## 🎬 Celeb DF(V2) Video Evaluation
-
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
+### Celeb-DF-v2
  
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_vit_b0** | 20 | conf | 0 | 0.9742 | 0.9877 | 0.0625 |
 | **ms_eff_vit_b5** | 20 | conf | 0 | 0.9900 | 0.9900 | 0.0408 |
 | **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.9842 | 0.9965 | 0.0283 |
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.9981** | **0.9984** | **0.0089** |
 
-</details>
+### FaceForensics++ 
+ 
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
+| **ms_eff_vit_b0** | 20 | conf | 0 | 0.9517 | 0.9860 | 0.1334 |
+| **ms_eff_vit_b5** | 20 | conf | 0 | 0.9842 | **0.9977** | **0.0477** |
+| **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.9808 | 0.9969 | 0.0637 |
+| **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.9850** | 0.9974 | 0.0492 |
+
+### KoDF 
+ 
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
+| **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.9655 | 0.9792 | 0.1237 |
+| **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.9792** | **0.9831** | **0.0692** |
 
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">🔗 Cross Evaluation</span></summary>
 
-#### 📌 Dataset: `celeb_df_v2` | 🧠 Model: `ff++`
+# Cross-domain Evaluation
 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+Each model is trained on one dataset and evaluated on a different, unseen dataset. Each table is titled `Train → Test`: the dataset the weights were trained on, followed by the dataset used for evaluation.
+
+### Celeb-DF-v2
+
+### Train: `ff++` → Test: `celeb_df_v2`
+
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_vit_b0** | 20 | conf | 0 | 0.7317 | 0.6949 | 0.6816 |
 | **ms_eff_vit_b5** | 20 | conf | 0 | 0.7143 | 0.7266 | 0.7662 |
@@ -94,34 +109,18 @@ To evaluate the generalization and robustness of our deepfake detection model, w
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.7722** | **0.7309** | **0.6657** |
 
 
-#### 📌 Dataset: `celeb_df_v2` | 🧠 Model: `kodf`
+### Train: `kodf` → Test: `celeb_df_v2`
 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.5579 | 0.4719 | 1.2605 |
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.5946** | **0.5400** | **1.0078** |
-</details>
-
-## 🎬 FaceForensics++ Video Evaluation
-
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
- 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
-| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
-| **ms_eff_vit_b0** | 20 | conf | 0 | 0.9517 | 0.9860 | 0.1334 |
-| **ms_eff_vit_b5** | 20 | conf | 0 | 0.9842 | **0.9977** | **0.0477** |
-| **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.9808 | 0.9969 | 0.0637 |
-| **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.9850** | 0.9974 | 0.0492 |
 
-</details>
+### FaceForensics++
 
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">🔗 Cross Evaluation</span></summary>
+### Train: `celeb_df_v2` → Test: `ff++`
 
-#### 📌 Dataset: `ff++` | 🧠 Model: `celeb_df_v2`
-
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_vit_b0** | 20 | conf | 0 | 0.5241 | 0.6860 | 1.1241 |
 | **ms_eff_vit_b5** | 20 | conf | 0 | 0.5533 | 0.7299 | 0.9987 |
@@ -129,94 +128,45 @@ To evaluate the generalization and robustness of our deepfake detection model, w
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.5825** | **0.7307** | **0.8897** |
 
 
-#### 📌 Dataset: `ff++` | 🧠 Model: `kodf`
+### Train: `kodf` → Test: `ff++`
 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_gcvit_b0** | 20 | conf | 0 | **0.4875** | 0.5341 | 1.6178 |
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | 0.4525 | **0.5902** | **1.5321** |
-</details>
-
-## 🎬 Kodf Video Evaluation 
 
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
- 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
-| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
-| **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.9655 | 0.9792 | 0.1237 |
-| **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.9792** | **0.9831** | **0.0692** |
+### KoDF
 
-</details>
+### Train: `celeb_df_v2` → Test: `kodf`
 
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">🔗 Cross Evaluation</span></summary>
-
-#### 📌 Dataset: `kodf` | 🧠 Model: `celeb_df_v2`
-
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_gcvit_b0** | 20 | conf | 0 | **0.4935** | **0.7258** | **1.3459** |
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | 0.4832 | 0.7160 | 1.4897 |
 
 
-#### 📌 Dataset: `kodf` | 🧠 Model: `ff++`
+### Train: `ff++` → Test: `kodf`
 
-| 🤖 Model | 🖼️ Frames | ⚙️ Agg Mode | 🔄 TTA (H-Flip) | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
+| Model | Frames | Agg Mode | TTA (H-Flip) | Test@Accuracy | Test@AUC | Test@Log Loss |
 | :--- | :---: | :---: | :---: | :---: | :---: | :---: |
 | **ms_eff_gcvit_b0** | 20 | conf | 0 | 0.7544 | 0.862 | 0.8829 |
 | **ms_eff_gcvit_b5** | 20 | conf | 0 | **0.7695** | **0.8821** | **0.7635** |
 
-</details>
-
-## 🧩 Branch-Ablation Study
-
-To quantify how much each branch — the **Low-level Branch** and the **High-level Branch** — contributes to MS-EffGCViT's performance, we trained and evaluated 3 variants (`Eff`, `Low-Eff-GCViT`, `High-Eff-GCViT`) in both `b0`/`b5` sizes across all three benchmark datasets.
-
-## 🎬 Celeb DF(V2) Ablation Evaluation
-
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
-
-| 🤖 Model | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
-| :--- | :---: | :---: | :---: |
-| EfficientNet B0 | 0.9643 | 0.9754 | 0.0977 |
-| EfficientNet B5 | 0.9768 | 0.9837 | 0.0676 |
-| low_eff_gcvit_b0 | 0.9788 | 0.9837 | 0.0658 |
-| low_eff_gcvit_b5 | 0.9865 | **0.9987** | 0.0150 |
-| high_eff_gcvit_b0 | 0.9807 | 0.9884 | 0.0343 |
-| **high_eff_gcvit_b5** | **0.9942** | 0.9977 | **0.0101** |
-
-</details>
-
-## 🎬 FaceForensics++ Ablation Evaluation
-
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
-
-| 🤖 Model | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
-| :--- | :---: | :---: | :---: |
-| EfficientNet B0 | 0.9202 | 0.9497 | 0.1523 |
-| EfficientNet B5 | 0.9660 | 0.9831 | 0.0842 |
-| low_eff_gcvit_b0 | 0.8033 | 0.9257 | 0.4148 |
-| low_eff_gcvit_b5 | 0.8700 | 0.9486 | 0.3365 |
-| high_eff_gcvit_b0 | 0.9783 | 0.9886 | 0.0767 |
-| **high_eff_gcvit_b5** | **0.9808** | **0.9956** | **0.0606** |
+# Inference Latency Comparison
 
-</details>
+Latency is measured with batch size 1 on a single Tesla T4 and video (num_frames=20).
 
-## 🎬 Kodf Ablation Evaluation 
+| Model | Mode | Params (M) | FLOPs (G) | Latency (s) | Ratio |
+| ----- | ---- | ---------- | --------- | -------- | ----- |
+| **ms_eff_gcvit_b0** | Fast Mode | 8.70 | 0.85 | 0.051 | 1.00× |
+| **ms_eff_gcvit_b5** | Pro Mode | 50.34 | 13.43  | 0.479 | 9.41× |
 
-<details>
-<summary><span style="font-size: 1.25em; font-weight: bold;">✅ Test Data Evaluation</span></summary>
 
-| 🤖 Model | 🎯 Accuracy | 📈 AUC | 📉 Log Loss |
-| :--- | :---: | :---: | :---: |
-| EfficientNet B0 | 0.9180 | 0.9520 | 0.2010 |
-| EfficientNet B5 | 0.9601 | 0.9770 | 0.1120 |
-| low_eff_gcvit_b0 | 0.8684 | 0.9481 | 0.2843 |
-| low_eff_gcvit_b5 | 0.8808 | 0.9596 | 0.2381 |
-| high_eff_gcvit_b0 | 0.9462 | 0.9696 | 0.1761 |
-| **high_eff_gcvit_b5** | **0.9745** | **0.9911** | **0.0844** |
+| Model | Params (M) | FLOPs (G) | Latency (s) | Ratio |
+| ----- | ---------- | --------- | ----------- | ----- |
+| **ms_eff_gcvit_b0 (Ours)** | **8.7** | **0.85** | **0.051** | **1.0×** |
+| sia (ECCV 2022) | 17.57 | 2.05 | 0.094 | 1.8× |
+| core (CVPR 2022) | 20.81 | 8.40 | 0.153 | 3.0× |
+| iid (CVPR 2023) | 65.5 | 12.7 | 0.181 | 3.5× |
+| effort (ICML 2025) | 303.4 | 155.9 | 1.041 | 20.4× |
 
-</details>
\ No newline at end of file