BitrateCraft · Video pipeline · bitratecraft.org
Filters, Scaling and Cropping
Filters sit between decode and encode, and their execution order is fixed by the pipeline rather than by the order of the controls on screen. Cropping runs first, then scaling, then the temporal and spatial cleanup stages. That sequence explains why a careless crop or an unnecessary rescale can quietly degrade everything downstream, and why HandBrake lists the active chain in its activity log as an ordered set of named stages. This page walks the chain in execution order and documents the parameters that matter most.
The order of operations
Each stage hands frames to the next. Geometry filters change the picture; temporal filters look across frames; spatial filters look within a single frame. Changing geometry late in the chain forces every later filter to process a different amount of data, which is why the fixed order exists.
Cropping
Source material often carries thin black borders left by the mastering process, and encoders spend real bits reproducing that black. Automatic cropping measures the edges and trims them, usually within a pixel or two of the correct value. Manual values are available for the awkward cases: letterboxed masters with intentional mattes, or transfers where a single bright pixel defeats the detector.
- Keep the cropped dimensions even. Odd values force chroma subsampling to be padded, and a few playback paths handle that badly.
- Re-check the crop after changing a scale or an anamorphic mode, since the numbers are expressed in source pixels.
- Be sceptical of a crop that removes a large fraction of the frame; a narrow band is usually a border, a wide one is usually content.
Resolution and scaling modes
The dimensions control offers an automatic mode that preserves the source size and a manual mode that sets a width and derives the height, or the reverse. A third option, often labelled with the storage-versus-display distinction, keeps the coded resolution while letting the display aspect ratio describe the shape.
| Mode | Effect | When it helps |
|---|---|---|
| Automatic | Copies source dimensions after cropping | Default; no resampling cost, no added softness |
| Fixed width or height | Resamples the frame, preserving aspect | Targeting a specific device class or panel size |
| Storage size preserved | No resample, shape stated in metadata | Anamorphic sources where the pixels are not square |
Scaling costs sharpness: every resample is a compromise, and upscaling invents detail that was never there. Downscale only when the destination genuinely cannot use the extra pixels, and prefer a single resample over a sequence of them.
Anamorphic handling
Anamorphic sources store a non-square pixel grid and rely on metadata to stretch it back to the intended shape. Three strategies are available. Leaving the flag untouched preserves the original intent exactly and is almost always the correct choice for archival work. Encoding square pixels removes the ambiguity and simplifies playback, at the cost of a resample. Setting a custom display shape is the escape hatch for material whose metadata was wrong to begin with. Whichever route is taken, the display aspect ratio in the log should match the shape a viewer expects.
Deinterlace and detelecine
Interlaced source shows comb-like edges on motion, and a progressive encode of interlaced frames keeps that damage permanently. The deinterlace filters combine pairs of fields into a clean frame; the better ones interpolate in the direction of the edge and discard less detail. Detelecine targets a different fault: footage shot on film and carried on an interlaced signal at a mismatched rate contains repeated fields, and the filter removes them to restore a true progressive cadence. Running both when only one is needed wastes time and can soften the picture, so check the source before selecting either.
Denoise and sharpen
Denoise suppresses noise that would otherwise consume bitrate. The lighter algorithm filters each frame independently and is quick; the stronger one compares a small neighbourhood of frames and preserves edges better, at a considerable cost in time. Both accept a strength value, and a modest setting usually beats an aggressive one. Sharpen sits at the opposite end of the chain and can partially offset resampling softness, but it also amplifies whatever noise survived. A common mistake is to pair heavy denoise with heavy sharpen, which removes texture and then re-invents it as ringing.
Filter decisions are cheap to test and expensive to reverse. Encode two short segments with and without the questionable filter, compare them at full size, and keep the simpler chain when the difference is not visible.
Common questions
Should cropping be manual? Rarely. Automatic detection is accurate for ordinary borders, and manual values are a fix for specific faults.
Does scaling improve quality on a small screen? No. It saves bits, and the picture is softer at every size.
Can filters be reordered? The chain order is fixed by the pipeline; the controls select which stages are active, not where they sit.
