deep-learning/image-forgery-detector-v2

Update app.py from main

Add model comparison script

Enlarge tab caption font size

Switch to complete dark professional theme

Fix header/footer consistency — full-bleed banner, unified page background

Add Model & Training tab with professional header/footer

Add CASIA scope note to UI

Add debug instrumentation (build version, input order, tensor stats, raw pred)

Fix ELA preprocessing to match training notebook (amplify + JPEG + tf.image decode)

Fix ELA computation to match train.py (ImageChops.multiply, not Brightness)

Fix: update sdk_version to 1.40.2, remove streamlit from requirements

Reference M3_best_v2.h5 (full model saved from TF 2.20)

Upgrade to Streamlit 1.40 + TF 2.20 (resolves protobuf conflict); fix use_container_width

Revert to TF 2.15 (protobuf-compatible with Streamlit); use cpu build

Load full H5 model via load_model; pin TF 2.20 to match training runtime

Fix build_model to exactly match train.py architecture and preprocessing

Fix missing layers import in build_model

Clean deployment: model loaded from HF Hub at runtime