1 The idea
CopyCat isn’t free. You hand-fix frames for the dataset, set it up, wait for training, sometimes retrain, QC every frame and fix the misses. It pays off on long shots with the same repetitive fix, and not on short shots or when every frame is different.
There are no ready-made numbers here: every value is your own estimate. Put in numbers from a real job and watch where the break-even point moves.
hand = N·h · ML = k·h + setup + (N − k)·(QC + r·h)2 Hands-on
Task: find the shot length where both routes cost the same artist time (within 3%)
Machine time (training + inference) isn’t artist time, but it uses the GPU and the schedule; count it too
3 Deep dive
Count all the time
The ML route isn’t just “waiting for training”. The real artist time goes into Ground Truth, building the dataset, retraining when it fails, QC on every frame and fixing the misses. Forget these and ML always looks better than it is.
| Part | Artist or machine time |
|---|---|
| Hand-fixing for Ground Truth | Artist |
| Setup, dataset, retraining | Artist (partly) |
| Training | Machine |
| Inference over the shot | Machine |
| QC and fixing misses | Artist |
The break-even point
The hand route grows with frame count at a slope of the minutes per frame. The ML route has a start-up cost (Ground Truth + setup) and then grows more slowly (QC + the share needing touch-ups). Where the lines cross is the break-even; longer shots favour ML.
If the time left per frame after ML (QC + touch-ups) is equal to or more than doing it by hand, the lines never cross and ML can’t win.
The more repetitive the fix, the lower r and the sooner it breaks even
What the numbers don’t say
Risk: if the model still fails after two training rounds, the time may be lost. Consistency: model output is often steadier than hundreds of hand-painted frames. Reuse: if a show has the same problem across many shots, a trained model can be reused or fine-tuned, which improves the sums further.
4 In real work
In Nuke
- Test before deciding: hand-fix 3–5 frames, run a short CopyCat training, check Inference on other frames, and estimate r from what you see.
- Look in the Cattery first: if a ready-made model does the job, you may not need to train at all.
- Time your real per-frame RotoPaint fix on the example frames so h is measured, not guessed.
Common mistakes → the fix
- ✗ Counting only the training wait✓ Count Ground Truth, setup, retraining, QC and fixing misses too.
- ✗ Using ML on a short shot because it feels modern✓ If the examples are nearly the whole shot, hand work or classic comp tools are faster.
- ✗ Borrowing someone else’s estimates✓ Measure on your own shot: run a small test first, then fill in the numbers.
Before you send the shot
- h (minutes per frame) has been timed for real
- r estimated from a short trial training
- Cattery checked for a ready-made model
- Machine time (training + inference) fits the schedule
- There is a fallback plan if the model fails
5 Check yourself
Which job is the best candidate for CopyCat?
🔒 Answer the question above and this card opens
1. Hand work is 4 min/frame. With ML you still spend 0.5 min QC plus touch-ups on 25% of frames. What is the ML time left per frame?
2. When can ML never pay off, however long the shot?
Where the numbers come from: Our simplified formula assumes the hand-fixed training frames are delivered as they are, and every frame that needs a touch-up costs a full hand frame. The defaults are examples to move, not industry figures. Nothing is trained in this lesson.