Spend Expensive Inference Selectively: Ranges, Auto-Split, and Special Media

Jasna's resource model becomes most useful when it controls not only how a job runs, but where expensive inference is allowed to run. This chapter examines selected video ranges, keyframe-aware smart rendering, LTX auto-split, VR180, and still images. The goal is to choose a full-video, scoped, mixed-model, or geometry-specific workflow without assuming that all of these paths preserve the same data or support the same restoration engine.

The common principle is selective change, but the selection can happen at different levels. A mask selects pixels inside a frame. A track selects related frames. A range selects scenes in time. An auto-split threshold selects which mosaics justify LTX. VR projection selects a coordinate system for each eye. A still image removes the temporal dimension altogether. Each choice changes the evidence and the boundaries that the rest of the pipeline must respect.

Scope video in time without pretending it is byte-identical

The Segment Editor and --segments let a reader choose only the time ranges that need restoration. Jasna still produces a full output video: selected spans are rendered, while the rest can be copied through the smart-render path. This is valuable for two reasons. It avoids spending restoration work on scenes without mosaics, and it keeps untouched spans closer to their source quality than a full re-encode would.

“Untouched” has a precise meaning here. It means the span is not sent through the restoration model and is usually stream-copied; it does not mean every packet around a boundary remains byte-for-byte identical. A video container can cut only at certain positions, often governed by keyframes. As a result, a few seconds around a selected range may be encoded again even when those frames are not themselves restored.

Jasna's smart-render path plans those boundaries around the input's keyframes. It creates copy fragments for spans outside the selected ranges, renders the affected spans, normalizes the fragments, concatenates them, and muxes the resulting full video. The output codec and frame timing must remain compatible with those operations. On one documented AMD H.264 case, the implementation falls back to a full re-encode when the source's B-frame structure cannot be stitched safely with the smart fragments.

This is a useful example of two different meanings of preservation. At the content level, the reader can preserve the original video outside the chosen scenes. At the packet level, the container may still require local re-encoding to create a valid continuous file. A scoped workflow therefore reduces unnecessary model work without promising that every byte outside the target remains unchanged.

The limits should be checked before a long job. Selected ranges operate on one video at a time, not folders or images. They require supported codecs and a constant frame rate, and they cannot be combined with streaming, high-FPS retargeting, fragmented MP4, or LTX auto-split. These are not arbitrary interface inconveniences.