Remove Document Shadow (DocShadow)

How it fits

Typical backends

Caveats

• LatencyThis is a learned restorer, heavier than a classical filter; CPU inference is slow for high-resolution pages, so a GPU is strongly preferred for live use.
• LatencyVery large pages are downscaled for processing to bound memory, then the result is resized back to the original resolution.
• CompatibilityThe checkpoint is fetched once from weights_url into the node cache; the first run needs outbound network unless the cache is pre-seeded.
• AccuracyThe model is trained on document pages; non-document inputs may be altered in unintended ways, so gate mixed streams upstream when needed.
• FallbackVery small inputs pass through unchanged, and device falls back to CPU when no GPU backend is visible.
• State lifetimedevice and weights_url are read once at startup; the model is loaded once and reused for every call.