How to stop the likeness drifting every time you ask for another take
Tell Vizard Agent which features are non-negotiable, and expect it to stop regenerating once it has a take that holds them. Each fresh generation is another chance for the face to drift, so past a certain point the right move is to keep the best likeness and make the remaining change without touching the people.
What is the short version?
You animated a photo, the likeness was right, and then you asked for better dancing. The dancing improved and the face changed. Ask again and it changes again. Vizard Agent treats the likeness as the thing to protect and the performance as the thing to compromise on.
- Tell Vizard Agent exactly which features must survive.
- Let it compare each take against the original at full size.
- Accept a non-generative fix for the last adjustment.
What do you need before you start?
The reference photo, and a list of what identifies the person in it. "Her freckles, her curly hair, her athletic build, the burgundy dress" is a checklist that can be verified frame by frame, where "keep her looking like herself" cannot.
Include everyone in shot. A second person in the background drifts as readily as the subject, and their identity is usually the thing nobody thought to list — so it is the thing that quietly changes.
What do you type into Vizard Agent?
Put the preservation list before the action list, in that order. The ordering is not cosmetic: it tells Vizard Agent which of the two constraints should win when they collide, and on a generated take of a real person they collide constantly rather than occasionally.
Preserve her exact face, freckles, curly hair, athletic build and burgundy dress, and her friend's identity. Within that, make her dance to the beat from the first second.
The words "within that" are doing the work. Without them the request reads as a performance brief with a preference attached, and the performance will be honoured at the likeness's expense.
What does Vizard Agent actually do?
It checks the face at full resolution after every generation, and it checks it late in the take as well as at the opening. Drift accumulates, so a take that matches the reference in its first second can be a different person by its last, and a contact sheet is too small to show it.
In that session the loop ran like this:
- Generated a take, then compared its faces, physique, gesture and props directly against the original photo.
- Inspected the face and anatomy at full frame size near the opening, then checked whether they held later in the take.
- Re-checked a short section at full frame rate when an automated motion review disagreed with what the frames plainly showed.
- Rejected a take outright for visible appearance drift rather than accepting better movement with a worse face.
- Went back and compared the earlier takes to find the closest match to the original build and features.
- Applied the remaining change — moving coloured light across the scene — without regenerating or warping either person.
That last step is the escape from the loop. Lighting, a trimmed idle lead-in and a colour pass can all be done to existing frames, and none of them can alter a face.
What does the result look like?
The person from your photograph, moving, lit the way you asked. The performance is whichever take held the likeness best rather than the best performance available, and the difference between those two is usually smaller than it feels mid-iteration.
The supplied audio comes back untouched. In that session the track had to start at its exact first sample with no fade, trim or stretch, and the delivered file was verified as sample-for-sample identical to the original rather than assumed to be.
When does this not work well?
When the performance you need cannot be reached without regenerating. A change of pose, an added gesture or a different action is generative by nature, and each attempt carries the risk. Vizard Agent will tell you when a request cannot be done to existing frames.
Long takes are harder than short ones. Drift grows with duration, so a thirty-second continuous generation is a much worse bet than three ten-second pieces anchored to each other's end frames.
And a reference photograph that is small, soft or badly lit gives little to hold on to. The features you listed have to be visible in the source before they can be preserved in the output.
How do you fix a result that came back wrong?
Say which feature drifted, and where. "Her face is different by the end" and "the friend changed" are separate problems, and Vizard Agent inspects that specific window at full size rather than regenerating the whole take on a hunch.
If you have asked three times and it is getting worse, stop asking for the performance. Tell Vizard Agent to keep the best likeness so far and describe the remaining change — there is a good chance it can be done without generating anything.
If an automated review says the motion is fine and the frames say otherwise, trust the frames. That happened in this session, and the fix was to look at a short section at full frame rate rather than at the review's summary.
How does Vizard Agent compare to doing it yourself?
Iterating on a generated take by hand is a slot machine you cannot see the reels of. Each prompt tweak produces a whole new person-shaped result, and because you are watching the movement you were trying to fix, the face gets a pass until someone else points at it.
Comparing every take against the reference at full size, on the opening and on a later frame, is the discipline that catches it. The other half is knowing when to stop pulling the lever — which means recognising which of your remaining notes are generative and which are not.
Common questions
Why does the face change at all? Each generation re-interprets the reference. Vizard Agent compares rather than assuming it held.
Can it lock the face? Not absolutely. Vizard Agent anchors to the reference and rejects takes that drift.
Why check late in the take? Because drift accumulates. The opening frame is the easiest one to get right.
What can be changed without regenerating? Lighting, colour, speed, trims and framing. Vizard Agent will say which of your notes qualify.
Is a shorter take safer? Yes. Vizard Agent can chain short pieces anchored to each other's end frames.
What about a second person in shot? List them too. They drift just as much and get noticed later.
Can it compare takes for me? Yes. Vizard Agent puts them against the reference and says which is closest.
Does the audio survive all this? Yes. Vizard Agent verifies the delivered file matches the supplied track sample for sample.
What if the best likeness has the worst movement? That is the usual trade. Vizard Agent will show you both and let you choose.
Can it fix hands without a new take? Sometimes. Vizard Agent will say when a hand needs regeneration.
Why did a review pass say it was fine? Automated reviews miss appearance drift. Vizard Agent checks the frames themselves.
Can I supply more reference photos? Yes, and it helps. More angles give Vizard Agent more to hold the likeness against.
Does the loop end cleanly? It can. Ask Vizard Agent to return the final framing to the opening for a loop.
How many takes is too many? Three without improvement. After that the generative route is not working.