This is how FullField Stitch actually works — the rig, the recording settings that make stitching easy, the seven-step wizard, and an honest look at the seam. Works with DJI, GoPro, Insta360, or any two matching action cameras.
An actual export — one continuous panorama assembled from two overlapping recordings of the same field.
The reference setup this walkthrough was shot on — deliberately unglamorous, because the software is what does the heavy lifting. Any pair of matching action cameras on a tall mount works the same way.
Two cameras never see the scene from exactly the same point, and the gap between the lenses is what creates parallax — the one seam error software cannot fully remove (measured below, honestly). Every centimeter you close between the lenses shrinks it at the source.
That’s the entire reason one camera rides upside down: it lets the two lens centers nearly touch. Everything else on this page is easier because of that one mounting decision.
Ten minutes of camera setup buys you a clean stitch. This checklist isn’t folklore — each item was tested on real footage, and the figures quoted are what was actually measured.
The app walks you through the whole pipeline in order. These are screenshots of the current build, not mockups.
Pick the two camera files. The metadata panel confirms both clips match — resolution, frame rate, codec — and a colour-profile readout flags flat footage (“Source looks flat — Flat → Rec.709 recommended”) so the right LUT is queued up before you ever see a washed-out export. If one camera rode upside down, tick its flip flag here.
The app finds your clap in both audio tracks and computes the offset between the recordings. For clock-synced rigs it also shows the timecode delta, with an agreement readout in frames — when the clap and the clocks agree, you know the sync is right.
A frame-level nudge for the rare case the automatic offset needs adjusting. Step either clip forward or back one frame at a time while watching a moving subject — when it stops doubling, you’re synced. Most sessions skip straight past this page.
Tell the app what lens shot each clip. Presets cover common camera models, so this is usually one dropdown per camera; the profile describes how the lens projects the world, which the stitch needs to undo before aligning the two views.
Scrub to a frame with action in the overlap zone — players crossing midfield are perfect. This is the frame the alignment solver will study, and texture in the overlap gives it far more to lock onto than empty grass.
Alignment computes automatically the moment you arrive — focal auto-tune, per-camera lens variance, and per-row correction all run without a button press, and a PASS/FAIL strip tells you whether the result is trustworthy. Drag the Level tool along the goal bases to square the horizon, then set the seam and crop. Good looks like: PASS on the strip and no doubled players at the seam.
The full render: pick the encoder (NVENC on NVIDIA cards), confirm the auto-recommended LUT, choose how many parallel workers to run, and go. Every export also appends a record to your local benchmark log — the same numbers behind the benchmarks page.
The same 30 seconds, three ways: the two camera files in, the stitched panorama out.
Left camera — covers the left half of the field. This is the file straight off the camera: 10-bit, flat log-style profile, washed out on purpose.
Right camera — covers the right half of the field, in the same flat profile. The color comes back later, in the export.
The stitched result — the same 40 seconds. About half of each camera’s view overlaps the other’s; the app aligns the two and blends them along a single seam, and the built-in LUT grades the flat log sources back to natural color in the same export pass — two washed-out half-field clips in, one continuous graded panorama out. These are sync-verified excerpts from a real session, downscaled only for this page.
Every two-camera stitch has a seam. Most products crop past it in their marketing; here’s ours up close, with the measurements that shaped how the app handles it.
The “5 px” you’ll see in the app is a feather width — how wide the strip is where the two cameras mix — not an alignment tolerance. A wide feather sounds safer, but it visibly ghosts anything that moves: at 200 px, tent poles and people double up. At 5 px the mix zone is narrower than the residual misalignment, so moving subjects stay sharp.
5 px feather — the seam strip is too narrow to double anything.
200 px feather — the wide mix zone ghosts tent poles and people.
Two same-model bodies on identical settings measured a ~2.6% effective scale difference — the same players, at the same instant, are visibly larger in one camera. You’d never notice it in either camera alone; at the seam it’s fatal. The app now solves this automatically at calibration: a shared-focal auto-tune, then a per-camera focal ratio, then a per-row correction.
Same players, same instant, both cameras — one image is measurably larger. This is lens variance between two identical-model bodies.
With that correction in place, seam misalignment on real footage went from 11.4 px to 1.1–2.3 px — measured solver-blind, on held-out frames the solver never saw. On the pitch itself that is now hard to find: in the clip below the halfway line runs straight through the crossover and players cross it without doubling.
What remains is depth. Two lenses a few centimeters apart see close and distant objects from different angles, and a single alignment cannot satisfy every depth at once — so the app fits the one that matters, the ground plane you actually film. Look at the mountain ridge along the top of the clip and you can still catch a small step where the two views meet: that is the residual, parked on scenery kilometres away instead of on the players.
The seam, unhidden — a real 30-second clip centered on the crossover, distant ridge to near touchline. The field lines and the players hold; the small step left in the far ridge is parallax physics, not a bug we’re hiding.
The panorama isn’t the video you share — it’s the source you cut everything else from.
Render at native resolution and maximum quality — every pixel you discard is ball-detection recall you can’t get back. The tracker reads the master once; give it everything the cameras captured. A sample master from two 4K clips measured 4504×2032 HEVC at 45 Mbps.
Pick whatever size suits your screen and bandwidth, then use the built-in Crop tool to frame the part of the field you care about. Because you cropped from a full-quality master, you can always come back and cut a different view — nothing was thrown away at export time.
The beta is free, everything runs on your machine, and the performance numbers are public.