Re-doing a shot without breaking the series
A regenerated shot is produced against whatever the model, the prompt and the references look like today. If any of those moved, the new shot can be better on its own and wrong in sequence, and nothing flags it. As of 2026-09-12.
| Input | Why it drifts |
|---|---|
| Model version | Updated by the vendor without notice |
| Prompt template | Improved by a colleague in the meantime |
| Reference set | Replaced with a better image |
Inclusion rule. Inputs that change silently between two generations of the same shot. Order. Alphabetical by input.
1The three things that move under you
The model version, the prompt template, and the reference set. Any of them changing between the original generation and the regeneration produces output that belongs to a different moment.
None of the three announces itself. A vendor can update a model without a note, a colleague can improve a prompt, and a reference image can be replaced with a better one.
2Check against the log, not against the shot
The temptation is to judge a regenerated shot on its own: does it look good. The question that matters is whether it matches the shots it will sit between, which is answerable only against the record of what those looked like.
This is why the continuity log names the file that established each character and location. Without it the comparison is somebody watching an old episode and trusting their memory.
3Regenerate with the original inputs, not the current ones
Keep the prompt, the seed where available, and the exact reference images used the first time, stored with the shot. Regenerating from the current versions of all three is the common path and it is the one that drifts.
Where the platform pins model versions, pin them. Where it does not, record which version was in force.
4Batch regenerations rather than scattering them
Redoing one shot in episode two today and another next month produces two different flavours of drift in the same season. Collecting the fixes and running them together, against one set of inputs, at least makes the drift uniform.
Uniform drift is far easier to live with than scattered drift, because an audience reads consistency as intent.
5Watch the neighbours, not the shot
After any regeneration, watch the sequence either side of it rather than the shot alone. Most continuity problems are only visible as a cut: the face is fine, the jacket is fine, and together they are not the same person walking through the same door.
Thirty seconds of viewing catches almost everything that a frame-by-frame inspection misses.
6Treat a model update as a reason to stop
If a vendor updates the model mid-series, the safe response is to finish the current episode on what you have and decide deliberately whether to move. Half an episode generated either side of an update is the worst of both.
This is also the moment to re-measure your attempt ratio, because a new model changes it in either direction.
7Keep the old version of a replaced shot
Do not overwrite. Keep the original alongside the replacement until the episode is delivered and signed off, because the judgement that the new one is better is often reversed once it is seen in sequence.
Storage is cheap and regenerating a third time is not.
8What this adds up to
Regeneration is not a repeat of a previous action. It is a new generation in a new context, and treating it as a repeat is how a season ends up looking like two seasons.
The habits above are unglamorous and between them they prevent the most expensive failure available in serial generated work: discovering in the edit that episodes one to three no longer match four to six.
Craft notes, not a register entry. No line here is attributed to a tool, and none should be read as a claim about one. The sourced material is on the register. Related: Keeping a face, Continuity log.
9Continuity in practice
Serial work like “Regenerating safely” depends on continuity; SceneMixer's guides describe how it keeps it:
The same way as a photographic one: the character is generated once as a reference sheet and that image is sent with every shot they are in, along with a couple of face-distinguishing notes. Shots refer to the character rather than re-describing their look.
SceneMixer, AI Animated Series Generator — “How do I keep a stylised character on model across episodes?” (read 2026-10-13)
It is the reference every later segment points at, which is what keeps a character's face and a location consistent across an episode.
SceneMixer, Glossary — “What is a setup image for?” (read 2026-10-13)