Appearance
Core concepts
Four ideas explain almost everything about how Kapnetix Post behaves. Read these once and the rest of the product follows. Throughout, a take is one recorded motion-capture performance, and a region is one span of frames on one marker inside a take.
Confidence is fixability
A confidence score predicts how cleanly Kapnetix can fix a take or a region. It does not merely say that something is wrong. This is the single most important thing to internalise: confidence is never an error meter.
You see the same confidence map at two moments, and they are deliberately symmetric.
- Predicted, before the clean. On the Diagnose screen, which is where you read what the diagnosis found, confidence forecasts how much review effort to expect, so you can plan the work before spending a token. A diagnosis is the read-only pass Post makes over a take when you upload it.
- Measured, after the clean. On the Review screen, which is where you read a finished cleanup, confidence grades the result Kapnetix actually produced, so you can verify it.
If a result cannot be measured, it is shown as not verified in the review queue rather than given a fabricated score. The banner on the Review screen says which of five things the measurement was. It is one of Verified, Assessed with exclusions, Limited assessment, Not repairable or Not assessed, and the measured drivers are listed underneath it. Only Verified is a clean bill of health.
What the take-level figure counts
A take has one confidence figure before the clean and one after it. Both count hand work in marker-seconds. One marker-second is one marker that needs an artist for one second. The fewer marker-seconds a take needs, the higher the figure.
- About 1 marker-second of hand work reads about 80, the lowest figure that is still OK.
- About 3 marker-seconds reads about 55, the lowest figure that still counts as Review.
- About 5 marker-seconds reads 37, which is Manual.
- Beyond that the figure falls quickly towards 1.
The figure before the clean is a forecast. Post adds up the frames of every region the diagnosis found and weights each one by how hard its repair is expected to be. A long occlusion weighs a lot, because the AI has to reconstruct several seconds of motion that the cameras never saw. The forecast also counts frames that are missing in the capture but that the diagnosis did not flag, for example on a performer it did not recognise. Each of those counts as one frame of hand work, the same as it counts after the clean, so a missing frame that the clean does not touch changes neither figure.
The figure after the clean counts what needs a hand. Post reads what the clean actually delivered, region by region. A region that shipped repaired adds nothing, unless it is a long fill. A marker swap that shipped repaired counts as repaired on both markers of the pair, over every frame where their labels were exchanged. These add hand work:
- frames that still have no data in the delivered file,
- regions the clean left alone, filled only in part, or handed back,
- joins between a fill and the captured motion that are still rougher than the limit,
- regions the diagnosis found that the run did not cover, for example because they were outside the frame range you chose.
A region that shipped open costs the measured figure exactly what it cost the forecast, so the measured figure never claims an improvement that did not happen.
A long fill still costs some hand work after it ships. A fill is the AI's reconstruction of motion the cameras never saw. When we compared our fills with the files artists finished by hand, the artists placed a growing share of each fill somewhere else as the fill got longer: little on a fill of a few frames, and more than half of the frames on a fill of a second or more. So the figure after the clean now charges every fill that is longer than about 6 frames at 120 frames per second (0.05 seconds). The charge grows with the fill's length, is higher when the check that compares the raw with the clean could not confirm the fill in the delivered file, and is higher again when a join at either end of the fill is still rougher than the limit. A fill never costs more than nine tenths of what the same gap would cost if it had been left open, so filling a gap always raises the figure. Short fills still cost nothing.
This is why a take can read low before and higher after, and why a take made of long fills does not read high after the clean. A take with many short gaps can read low before the clean and 99 after it once they ship filled. A take with several multi-second occlusions reads low before the clean and stays in Review or Manual after it, even when every occlusion ships filled, because those fills are the work an artist checks first. A high measured figure says that little is left to do. Each region card on the Review screen keeps its own confidence for the fill, and long fills are the ones to look at first.
Raw is immutable · Clean vN is non-destructive
Your raw capture is the file you uploaded, and it is never modified. Every cleanup run stacks beside it as a new named version, called Clean v1, then Clean v2, and so on. The version history is never collapsed. You can always compare a clean version against the raw, or against an earlier clean, and you can revert individual regions.
A run is a versioned batch of takes cleaned together. A single take's history shows the run versions that produced it.
Quality bands and take states
Two questions are answered the same way on every screen. Can I trust it? is a coloured quality badge with a word, a shape and a score from 1 to 99. Where is the take? is a plain state chip, such as Raw or Cleaned v2, which never carries a quality colour. The badge says predicted on the Diagnose screen, before a take is cleaned, and measured on the Review screen, after it.
| Band | Score | Shape and colour | What to expect |
|---|---|---|---|
| OK | 80 and up | green circle | No review needed, or a quick look is enough |
| Review | 55 to 79 | amber square | Needs your eyes |
| Manual | under 55 | red triangle | AI assists · you finish |
Here is what the bands mean.
- OK means Kapnetix expects to fix the take, or has fixed it, with little or nothing left for you. A take that arrived clean reads Raw with an OK badge and good as captured: there is nothing to fix, and it is kept as raw.
- Review means the take was or will be cleaned, but it has borderline regions worth a look.
- Manual means it could not be cleaned by the AI alone. The AI assists and you finish the pass.
A region or a marker shows its band without a second score. Only a take that is good as captured stays raw. Every other take is listed under To clean.
On the Review screen, the queue lists every cleaned take with its state chip and its badge. A take with any region that needs eyes, for example a spike left in the file or a gap that was not rebuilt, never reads No review needed. Its status says how many regions need eyes, for example 3 regions that need eyes, it sits with the takes that need review at the top of the queue, and it counts in the takes need eyes figure in the screen header. No review needed is shown only for a take with no regions left that need eyes. While Post is still counting a take's regions, its status reads counting regions. If the regions could not be counted, the take reads Needs your eyes and counts among the takes that need eyes.
The verdict follows how much of the take was measured against the raw capture. With nothing open and every repaired frame measured, the take is Verified. With nothing open but only part of it measured, for example an Assessed with exclusions or Limited assessment take, it never reads No review needed. It reads Verified where measured · review the rest by eye in the queue and on the Projects screen, and the verdict on the Review screen says the same.
Post uses one vocabulary for a region's state. An open region is a region that needs eyes, and a closed one is repaired. The word errors names only what the diagnosis detects, on the error matrix and the error list of the Diagnose screen.
A region needs eyes when its repair did not ship whole, when it was filled as a best attempt (including a gap the diagnosis did not flag), when a spike was left in the file, when a join where its fill meets the captured frames is still over the limit Kapnetix sets for a join, when it shipped with no score or a score of 0, when no repair ran on it, or when you reverted it to the raw capture. Two kinds of region are left as captured on purpose instead, and do not count as needing eyes: a jitter span too short for Post to smooth, and a region on markers or error types you left out of the cleanup. They are listed and counted in their own group on the Review screen, after Needs your eyes and Repaired and verified. The same count and the same badge appear beside each Clean version on the Projects screen, for example Clean v2 · 3 regions that need eyes. A version with none keeps its measured badge and reads No review needed.
How much hand work to expect
The Diagnose screen header shows the selected take's predicted quality badge, its cleanup forecast band and the hand work they imply, for example Manual 24 (predicted) · Manual · about 6 min of hand work. The line under it adds the figure up for the batch.
The hand-work figure is a planning estimate made from the diagnosis, before any cleanup runs. Post allows about 6 minutes for a take whose predicted band is Manual, 3 for Review, 1 for OK and none for a take that is good as captured. The batch figure adds up the takes in To clean whose diagnosis has finished. A take whose diagnosis failed or is still running is not counted.
Confidence heat vs. quality badges
The badges above are categorical, in that each one is a single choice out of three bands. Separately, the brand pink to cyan gradient is the continuous per-frame confidence scale on the timeline and the 3D viewer, where it also colours the markers. Cyan is the OK end and pink is the Manual end, on a 0 to 100 scale. The viewer's legend names the two ends with the band words. Badges stay categorical, and heat stays continuous.
Errors and the model that fixes each
Kapnetix classifies what it finds into five kinds, and each kind is repaired by a specific model.
| Error | Label on screen | What it is | Model |
|---|---|---|---|
| Gap | Gap / occlusion | Missing markers / occlusion | Solve · Gapfill |
| Merge | Marker merge | Two markers collapsed onto one | Solve · Gapfill |
| Swap | Marker swap | Marker labels swapped | Solve · Relabel |
| Jitter | Jitter | High-frequency noise | Solve · Denoise |
| Spike | Spike | A short, sharp impulsive peak | Solve · Denoise |
A class this build of the app cannot name shows as Other and claims no model, rather than being guessed at.
Each region belongs to a performer or to a prop, an object such as a box or a staff with markers of its own. A prop is repaired as one rigid piece, placed from its other markers, and its regions are listed under it with the same colours. See Props.
The order the work runs in
The principle does not change: complete the skeleton first, correct identities next, and treat noise last, once positions are right. A jitter cut runs after the fills of its own pass for the same reason. A cut has to rejoin the trajectory on both sides, and beside a hole there is nothing to rejoin to.
The exact order is worked out per take, from what the diagnosis found and what each step needs to be true first. That order is not hidden. It is the Cleanup Plan, the per-take list of steps Post will take, and you can read it before you launch. See The Cleanup Plan.
In Auto mode, which cleans every take in To clean unattended, Kapnetix picks the model for each region and plans the order for you. In Assisted mode you scope the work yourself by setting the frame range, the error types, the markers and the jitter sensitivity, and the plan is built inside that scope. The governing rule throughout is that the AI only touches what you scope.
What these ideas do not claim
- Confidence is not a promise. It is a prediction before the run and a measurement after it, and the two can disagree. Showing both is exactly the point.
- A take that is good as captured is not a cleaned take. The diagnosis found nothing in it, and it is kept as raw.
- A model name on a region card is not a guarantee that the fill shipped. A card that reads Nothing filled carries no model and no score, because nothing was fixed.
What to do next
Read how Post cleans a take for what happens between Diagnose and Review, or walk the quick start on a real file.
Ask us
Something looks wrong on a take? Write to support@kapnetix.ai and send the take id, or the take's name and the project it sits in.