Workflows

Advanced Tracking

What to reach for when a solve fails, drifts or comes back with the lens slightly wrong.

Basic Tracking covers the happy path. This page is for when it doesn't work, and for squeezing a better lens out of a shot that did.

The first half is a ladder. Work down it in order, because each step costs more time than the one above it.

1. Raise The Quality

Quality on the Auto Track node controls how many features are extracted per frame.

The symptom that calls for High is specific: the solve completes but only some of your frames come back registered. More features per frame means more chances to match across a difficult stretch. Complex scenes and zoom footage are the usual culprits.

It's slower, so don't start here. But if Medium half-solves a shot, High is the cheapest next thing to try.

2. Mask Moving Objects

Anything moving on its own generates features that contradict the camera motion. A person crossing frame, a passing car, leaves in wind, a reflection sliding across glass. The solver has no way to know those points aren't part of the world.

Add an AI Mask, prompt for whatever moves, connect it to Mask Frames, and leave Mask Mode on Exclude Masked Area. Dilate the mask a little so the subject's edge pixels go with it.

See AI Mask. If the moving thing is something you actually need tracked, see Object Tracking instead.

3. Turn On Loop Detection

Loop Detection spots when the camera comes back to somewhere it has already been, and adds matches between those distant frames.

That matters because sequential matching only compares neighbouring frames. Small errors accumulate across a long shot, so by the time you return to your starting point the solve has drifted and doesn't realize the two views are the same place. Loop detection ties them together and pulls the drift out.

Turn it on for shots that orbit a subject, walk a circuit, or pan away and come back. Leave it off otherwise, because it indexes every frame at roughly half a second to a second each, and a shot that never revisits anything gains nothing.

It needs SIFT features, so it's skipped on the LightGlue backend.

4. Switch To LightGlue

Features picks the backend that finds and matches points.

SIFT is the classic algorithm and the default. Fast, especially on GPU.

LightGlue (AI) uses learned features and learned matching instead. It's considerably more robust on the footage SIFT struggles with:

• Steep viewing angles, where a patch looks very different from frame to frame
• Repetitive texture, like brickwork, tiling or fencing, where SIFT matches confidently to the wrong brick
• Low-texture surfaces that barely give SIFT anything to hold

It needs the AI runtime installed, the same one AI Mask uses, and it wants a GPU. Loop detection doesn't run on this backend, so if you need both, stay on SIFT.

5. Add Manual Tracks With Enhance

When the solve mostly works but has gaps, or jumps at a particular moment, hand-tracked points fix what automation couldn't.

Track a few features through the problem area in the Movie Clip Editor, wire them into a Manual Tracks node, and feed that plus your solve into Enhance Auto Track.

Manual Track Weight decides how hard those points pull compared to automatic features. It defaults to 8, meaning your hand tracks count for eight times as much, which is usually what you want since you placed them deliberately.

Bridge Missing Frames handles gaps. Frames the Auto Track never registered get placeholder poses, interpolated inside a gap and extrapolated at the ends, then corrected by any manual tracks covering them. For a real recovery rather than a smooth guess, hand-track at least three points through the gap and keep tracking a little way into the solved section on both sides, so the solver can tie the recovered frames to something it already trusts.

Leave Confidence Report on. It tells you about parallax health, whether your scene is too planar to solve well, how much of the frame the lens estimate actually covers, and how certain each solved parameter is. That last part is worth reading before you trust a lens.

Lens Calibration

Everything above is about getting the camera solved. The rest is about getting the lens right, which matters as soon as you need CG to sit convincingly in the plate.

Grid Shots

The most reliable lens data doesn't come from the shot at all. It comes from photographing a checkerboard.

Shoot 15 to 30 stills of a printed checkerboard with the same lens at the same focal length, filling different parts of the frame and tilted at a range of angles. Corners matter most, because that's where distortion is strongest.

Point an Images Folder at them, run Lens Grid Calibrate, and connect its Lens Data output into Auto Track's Lens Data input.

Set Square Size (mm) to your printed square and Sensor Width (mm) to your camera's sensor. Board dimensions are detected automatically. If you know the lens focal length, type it into Focal (mm) to lock it and let the calibration solve only distortion.

Feeding known lens data into a solve removes a whole set of unknowns and usually improves everything downstream.

Plumb Lines

No grid shots, and a lens that bends visibly? Draw on the curves instead.

Enhance Auto Track has a Plumb Lines picker that takes a Blender mask. Draw splines along edges you know are straight in the real world, following the curve as it appears in the image. Building edges, door frames, road markings, poles. Three or more points per spline.

The solver knows those should be straight, so any curve is distortion, and that's strong evidence about the lens.

Set Lines Frame to the frame you drew on. If you want lines on several frames, name each mask layer with its frame number, like 36 or frame 120, and that layer pins itself there. Useful on a zoom, where you can constrain the lens at several focal lengths. Line Weight controls how hard they pull.

This is the practical answer for heavily distorted footage with no calibration data: solve normally, look for edges that should be straight and aren't, and draw on them.

Multi-Shot Lens

When several shots came off the same physical lens, solve the lens across all of them at once instead of separately per shot.

Multi-Shot Lens pools up to three solves of the same lens into one joint calibration. Connect each shot's Auto Track to a numbered input, and each comes back refined with the shared result.

More views of the same optics means a better-constrained answer than any single shot can give, particularly when each shot only covers part of the frame or has limited camera movement.

Blender's Lens Model Limitation

Dope Track solves rich lens models: Brown-Conrady with up to four radial terms, Kannala-Brandt for fisheye, Double Sphere for very wide lenses.

Blender's own movie clip camera can't hold any of that. When results are written back to the clip, only k1 and k2 survive, because that's all its tracking camera has room for.

So Blender's built-in distortion display and its own undistort preview are showing a simplified version of what was actually solved. On a mild lens the difference is small. On a wide or fisheye lens it can be badly misleading, and it will look wrong in ways that tempt you to go re-solve something that was already correct.

Judge distortion with the Lens Distortion node, which uses the full solved model. See Lens Distortion and Comp.

Lining Up Against Known Geometry

If you have a model of something in the shot, whether a surveyed set, a measured building or a CAD part, you can link hand-tracked markers to its vertices and solve directly from those pairs. It works either way round: hold the model still and solve the camera, or hold the camera still and place the model.

The same links can also calibrate the lens, since known 3D geometry seen through a distorting lens tells you plenty about the distortion.

Full detail in 2D/3D Constraints.

Settings Worth Knowing

The four at the bottom of the node are solver tuning. The defaults are sensible and most shots never need them touched:

Tightening the two loss thresholds can clean up a noisy solve at the cost of throwing away points. Loosening them keeps more points and more error with them.