Upscale Image (DRCT)

@ming053l1 VersionEnhance

Super-resolves an Image with the DRCT transformer at a fixed integer factor matched to the checkpoint, optionally tiling large frames to bound VRAM, and emits the enlarged Image.

How it fits

Typical backends

Recover detail in low-resolution captures and publish the enlarged image: a file source feeds frames, this component upsamples them, and the result is streamed out over HTTP.

Upscale small text crops before OCR so the recognizer sees larger glyphs, then publish the recognized text as JSON.

Caveats

Hard constraintThe upscale factor must match the loaded checkpoint family; a mismatch fails at load time rather than at inference.
Parameter interactionThe minimum-width and minimum-height floors gate the skip TOGETHER: upscaling is skipped only when every non-zero floor is already satisfied, so an image that already meets the size targets is passed through untouched. A floor of zero removes the constraint on that dimension.
Parameter interactionLarge frames are processed in overlapping tiles to bound VRAM, with the tile size trading memory against per-tile context and the overlap blended to hide seams; the output can also be clamped to maximum dimensions after upscaling.
LatencyThis is a heavy transformer super-resolver; a GPU is effectively required for live use and CPU inference is slow.
I/O contractThe first run downloads the checkpoint from the configured repository unless the cache is pre-seeded; the output is a same-color-space enlarged Image.

Versions

b9d520d1latestdefault
Automated release
6/24/2026