Why Clip Boundaries Matter: Stability, VRAM, and Seam Control
Once Jasna has turned frame-level detections into tracks, it still has to decide how much of each track to process at once. That decision is the clip boundary. This chapter follows a tracked region through maximum clip size, overlap, continuation, and blending, then connects those transitions to VRAM, repeated work, compilation, and recovery. The aim is practical: when a result has a seam or a job runs out of memory, the reader should be able to make one bounded tuning change instead of changing unrelated settings at random.
A clip is not just a convenient batch. It is the time window in which a temporal restoration model can use related frames together. If the window is too short, neighboring clips repeat boundary frames and may have less context for a stable result. If it is too long, the model and its intermediate data occupy more GPU memory. Jasna's design makes that trade visible rather than pretending that one fixed clip length fits every video and every card.
A tracked region has a lifecycle
When a detection continues to match an active track, the clip grows frame by frame. The tracker stores the region's boxes and masks, and the processing pipeline stores the corresponding crops for later restoration. This state persists until one of several transitions occurs: the region disappears for too long, a scene cut ends it, the configured maximum clip size is reached, or the input itself ends.
The maximum clip size is the longest clip that Jasna will process as one unit. Reaching that limit does not necessarily mean that the mosaic vanished. With temporal overlap enabled, the current clip ends and a continuation clip begins with shared frames from the end of the previous one. Those shared frames give neighboring clips common temporal material. The overlap is then used to make the join less abrupt and to let the final pipeline select the parts of each result that should be kept.
This is an important distinction between a boundary caused by content and a boundary caused by resource management. A scene cut says that two shots should not be mixed. A maximum clip size says that one continuous track should be divided into manageable pieces. The output should preserve that difference: the first transition protects identity across a change in source content, while the second tries to preserve continuity across an artificial processing limit.
The processing stage has to manage both sides of the split. It can discard parts of a clip that are only there to provide overlap, retain the central work that belongs in the output, and attach the continuation's shared frames to the next unit. If crossfade is enabled, the overlapping results can be weighted at their boundary. The system is not simply concatenating two independent videos; it is reconciling two restoration results that cover some of the same time.
Why bigger clips can be faster
The intuitive assumption is that smaller batches should always be faster because each one is easier to fit into memory. For Jasna's BasicVSR++ path, that assumption is often wrong. Every additional clip boundary can force overlap frames to be processed again.