Expand description
Tier: SwinRestore — SwinIR transformer SR (heavy, tiled for large stills).
Permissive weight source (gated ①, see docs/plans/gated-av-jobs-2026.md §2,
row “SR SwinIR-heavy”): SwinIR (MIT/Apache), converted to ONNX, dropped
into vendor/vision/sr/swinir/. Large stills are processed tiled by the
existing SR tiling runtime; this adapter runs one tile/frame per call.
Pipeline: SwinIR is a full RGB restorer. We normalise RGB8 → [0,1]
planar 1×3×H×W, run the ONNX model, and denormalise the
1×3×(H·s)×(W·s) output to RGB8 into out.
This adapter targets the restoration / Sharpen SwinIR variants and is
therefore reported non-generative (generative = false): it reconstructs
detail without hallucinating texture, keeping it usable for the honest
(medical/forensic) path. (SwinIR also has GAN “real-world” variants — those
would be wired as a distinct generative tier, not here.)
#[cfg(not(feature = "vision-onnx"))]→ fail closed (FeatureDisabled).#[cfg(feature = "vision-onnx")]→ realortinference.
No unwrap() outside #[cfg(test)].
Constants§
- BACKEND_
ID - Backend id reported for this tier.
Functions§
- swin_
super_ resolve - Super-resolve
rgbbyscaleusing a SwinIR (restoration) ONNX weight.