Overview

Jasna converts a detected box and mask into an enlarged context crop, normalizes it for BasicVSR++'s fixed 256-by-256 output, optionally enlarges the result, and composites it back through the mask. The design keeps source pixels outside the selected region, but crop scale, mask shape, VR projection, and secondary restoration affect the boundary. Enlargement can improve apparent fit or sharpness for large targets while adding time, VRAM, and possible flicker; it does not recover information that the primary inference lacked.

Learning outcomes

The 256-by-256 primary patch is a stable interface, but large targets and close-ups can look soft when resized back to the frame.
Secondary restoration changes patch scale and presentation; it adds cost and can amplify errors or introduce flicker.
Masks and VR-aware delta compositing limit where generated changes act, but they cannot verify the inferred content inside the selected region.

Keywords

JasnaCropsPatchesMasksCompositingBasicVSR++

The Smallest Useful Change: Restoring a Patch Without Rewriting the Frame

Jasna's most distinctive promise is spatial: restore the selected mosaic area while leaving the rest of the frame alone. That promise is implemented through a chain of crop preparation, fixed-size inference, optional enlargement, and masked compositing. This chapter follows one detected region through that chain so the reader can see both sides of the design: why locality protects untouched pixels, and why softness, mask artifacts, and uncertain detail can still remain inside the selected region.

The previous chapter treated time as the main boundary. A tracked region had to be divided into clips without losing continuity. Here the question is spatial: what exactly crosses from a frame into the restorer, and what comes back? The answer is not a replacement frame. It is a generated patch with a known relationship to a source crop and a defined place where it may influence the original image.

The crop supplies context to a fixed-size model

Jasna begins with a detected box and mask, but the box is not necessarily the exact crop sent to the model. The public crop preparation expands the region, subject to the frame and the relevant eye boundary for VR processing, so the restoration sees surrounding context. That context can help the model interpret edges, shapes, and local appearance instead of treating the mosaic as an isolated rectangle.

The expanded crop is then resized and padded to the input size expected by the primary BasicVSR++ restorer. The model returns a 256-by-256 patch. The fixed size creates a stable interface between the spatial preparation stage and the learned model: regardless of the original target's dimensions, the restorer receives a normalized patch shape and returns the same basic output shape.

That stability is useful, but it is not free. If the detected area is small, the crop can be represented at a scale that gives the model useful context. If the area is large or the video is a close-up, the same 256-by-256 output has to stand in for a larger portion of the frame. A patch can then look soft when it is resized back to its original region. The limitation is not necessarily that the detector failed; it can be the spatial contract of the primary restorer itself.

The crop also determines what the model does not see. Expanding the box provides context, but it does not turn the crop into a full-frame reconstruction. Information outside the crop is available to the blend stage as original pixels, not as additional generated evidence. If the selected area is too broad, the job spends more effort and risks changing more content. If it is too narrow, the model may have less surrounding structure to guide the patch and the boundary may be harder to integrate.

Secondary restoration changes scale, not history

Jasna offers optional secondary restoration for cases where the primary patch is too small for the target.

Published by DeepMox

FAQ

Why does Jasna use an expanded crop instead of the exact detection box?

Jasna expands the detected region to give the restorer surrounding context for edges, shapes, and local appearance, while the mask can still restrict where the generated patch acts.

When does Jasna need secondary restoration?

Jasna may benefit from secondary restoration when a large mosaic, close-up, or 4K target looks soft relative to the 256-by-256 primary patch. The added stage costs time and VRAM and can flicker on some videos.

How does Jasna keep VR180 patch placement local?

Jasna processes the left and right eyes separately and can project the restoration delta back into source space, reducing reprojection changes outside the target. This still depends on correct projection and per-eye detection.

Previous7 minJasna Clip Boundaries

Next7 minJasna Quality Chain

Jasna Regional Patch

The Smallest Useful Change: Restoring a Patch Without Rewriting the Frame

The crop supplies context to a fixed-size model

Secondary restoration changes scale, not history

FAQ