segAll tamper localiser (trained 2026-09-22)
Document-tamper localiser: given a rendered document page image, outputs a per-pixel
tamper-probability logit map. SegFormer-B2 (nvidia/mit-b2 backbone) widened from 3 to 6
input channels (RGB + 3 fixed SRM high-pass residual channels computed from luminance),
num_labels=1 binary tamper-vs-clean logit map.
Usage
from transformers import SegformerForSemanticSegmentation
model = SegformerForSemanticSegmentation.from_pretrained(
"mymani/segAll-tamper-localiser", revision="2026-09-22"
)
pixel_values must already be the 6-channel (RGB + 3 SRM residual) input described above —
see the reference wrapper linked above for the exact preprocessing.