INDEPENDENT RESEARCH / DOCUMENT INTEGRITY
A closer look at document tampering. TAFE-ID explores how visual and frequency evidence can reveal manipulated text, one region at a time.
REFERENCE MODEL / SOURCE PAPERToward real text manipulation detection: New dataset and new solution ↗Dongliang Luo et al. · Pattern Recognition, 2024 · RTM / ASCFormerREFERENCE INFERENCE VERIFIED / RESEARCH PROTOTYPE
Recorded ASCFormer results on a 32-document verification subset. These numbers are not a full benchmark or a claim about the custom TAFE model.
01 / THE WORKBENCH
A segmentation mask shows where the model suspects an edit. The workbench shows a saved ASCFormer self-test output. You can also preview your own document locally; new predictions require the linked Colab runtime.
Recorded self-test screenshot · edit_0192REAL MODEL OUTPUT / SAVED RUN
This is the saved output from our successful ASCFormer self-test on RTM edit_0192. The red region was produced by the reference model. Viewing it does not rerun inference, and it is not a ground-truth annotation.
Recorded run: 1122 × 794 pixels · predicted region 0.279% of pixels · maximum model probability 0.969. These are output diagnostics, not accuracy scores.
View image source and license ↗Run the verified GPU demo in Colab ↗New predictions require a running Colab session. This website currently shows a recorded result and local previews; it does not host the model.
02 / TWO RESEARCH TRACKS
The RTM authors’ released model is the working reference. Strict checkpoint loading, CUDA operators and document inference passed in the isolated environment.
Our experimental pipeline reads real JPEG coefficients and quantization tables alongside RGB crops. Learning fixed crops did not establish useful held-out performance.
RTM = REAL TEXT MANIPULATION, THE DOCUMENT DATASET USED IN THIS PROJECT.
03 / PERFORMANCE PROFILE
Reference-model F1 by manipulation type, from the same 32-document subset. Small category counts limit what we can conclude.
04 / LAB NOTES
9,000 RTM images and masks were restored. Crop alignment and actual JPEG coefficients were inspected before frequency experiments.
A balanced GroupNorm baseline passed its four-crop gate. Custom models remained weak on held-out documents. BatchNorm statistics and loss balance were investigated.
The authors’ checkpoint produced usable localization and a working Colab demo. Its results remain distinct from our custom training experiments.
05 / REPRODUCIBILITY
The verified package is saved in the project owner’s Drive as TAFE-ID-Final.zip. The GPU demo runs separately from this website; arbitrary image uploads are not connected here.
Download recorded verification JSON ↗