Unwarp Document (DvD)

@hanquansanren1 VersionEnhance

Rectifies a curved, folded, or crumpled document Image with the DvD coordinate-diffusion model, which runs an iterative denoising loop to predict a backward warp and emits the flattened Image for OCR.

How it fits

Typical backends

Flatten a heavily crumpled page before OCR: a file source feeds the image, this component runs the diffusion unwarper, and a recognizer reads the cleaned text out as JSON.

Rectify documents from a live HTTP feed and publish the flattened image for visual inspection instead of OCR.

Caveats

LatencyCoordinate-level diffusion runs an iterative denoising loop per frame, so it is markedly heavier than the single-pass rectifiers; a GPU is strongly recommended and CPU execution is impractical for live use.
Hard constraintThe checkpoints live in a gated repository, so an access token whose account has accepted the repo conditions must be supplied before the weights can be fetched on first run.
AccuracyThe backward map is denoised at a small resolution inside a bounded working frame and upsampled, so very high input resolutions do not recover finer detail; hands, torn edges, and non-document inputs remain common failure cases.
I/O contractAn image smaller than two pixels on a side is returned unchanged; otherwise the output is a same-color-space rectified Image.
LimitationThis is a dewarping-only restorer with no deshadow, deblur, or binarize modes; pair it with a dedicated cleanup stage if illumination correction is also required.

Versions

4938afabdefaultlatest
rebuild on GPU-detection-fixed runtime
6/17/2026