When the Workflow Breaks: Diagnosing Misses, Seams, and Resource Limits
A failed Jasna job rarely announces its cause in the same words as its symptom. An untouched mosaic may reflect detection rather than restoration. A seam may come from a dropout, a scene cut, a clip split, or a mask boundary. A long wait may be model preparation, a resource mismatch, or a delivery choice that cannot keep up. This chapter uses the pipeline's ownership boundaries to diagnose those cases. The method is deliberately modest: form a few plausible causes, run the least expensive test that distinguishes them, make one bounded change, and stop when the evidence points to an unsupported path.
The goal is not to memorize every setting. It is to avoid changing the wrong stage. A more detailed restorer cannot recover a region that never entered the pipeline, and a larger clip cannot make an incompatible container stitch safely. Diagnosis begins with the visible behavior, not with the most impressive option in the interface.
An untouched or unwanted region starts at selection
If a mosaic remains untouched, the first hypothesis should be a missed detection rather than a weak restorer. The restoration stage receives selected crops; it cannot act on a target that the detector did not identify. Jasna's detector guidance offers several ways to test that hypothesis: use the recommended model for the content, consider a larger model for 4K, use a VR-specific detector for VR180, or adjust the detection threshold. Lowering the threshold can help when mosaics are missed, while raising it can reduce false detections.
A short scan or preview is a better first test than a full LTX render. If the target never appears in the detected regions, changing the restoration engine is premature. If the target appears but the result is weak, the problem moves downstream to crop geometry, model choice, temporal evidence, or blending. The distinction saves time because it separates “no input to the restorer” from “an unsatisfactory output from the restorer.”
The opposite symptom is an unwanted patch on ordinary content. A one-frame or very short detection can be a false positive. Jasna's minimum detection duration can discard detections shorter than a configured number of frames; its documented default is two frames. But increasing that setting can also remove a genuine short mosaic. The test should therefore use the source's actual timing: if the target truly exists for a brief interval, a more aggressive filter solves one problem by creating another.
The maximum detection gap creates the corresponding diagnosis for intermittent targets. If a mosaic is visible, disappears briefly, and returns in the same place, bounded gap filling can preserve one continuous track. If the target moved or vanished, a longer gap can reuse an old box and mask over ordinary content. The gap control is a hypothesis about detector instability, not a license to treat missing evidence as present.
Detector selection itself remains conditional. The repository describes the default RF-DETR model as fast and suitable for most videos, while other models serve 4K, VR180, animation, or AMD speed. Those recommendations are starting points, not independent accuracy guarantees. If two detectors disagree, the result should be treated as evidence to inspect rather than as proof that one model is universally correct.