Your reference stops working by shot four
Chaining each clip to the last output compounds identity error. Re-anchor every new shot to the master reference, not to shot three's last frame.
Shot one is the person you approved. Shot two is close. Shot three is a cousin. Shot four is a recast. You did not change models and you did not get sloppy with the prompt. You fed each clip the last frame of the clip before it, and you taught the model that whatever it had slightly invented was now ground truth.
That is chaining used for the wrong job. Scene chaining is how you continue a motion past a duration ceiling: last frame of clip N becomes the first frame of clip N+1. It is the right tool for one action that does not fit in a few seconds. It is the wrong tool for a sequence of new shots, because each generation adds a little identity error and the next generation treats that error as the character.
The re-anchor rule is the other discipline: every new shot goes back to the master reference, the still you would reprint the character from, not to the last frame you happened to like. Motion continuity and identity continuity are two different inputs. Mixing them is why the lead has a new jaw by the middle of the scene.
Why shot four is a different person
Nothing carries between generations by default. Chaining is a way of carrying pixels instead of adjectives, which is why it works at all. The pixels it carries are the last frame of a generated clip. That frame is, structurally, the dirtiest frame in the clip: error accumulates along the duration, and the final frame is where that accumulation is highest. Cleaning the handoff frame exists because people hand that dirt forward unmodified.
Even a cleaned handoff is still a descendant of shot one, not a copy of the master. A slight yaw in the jaw, a collar that crept, a pupil that sat slightly wrong: each is small enough to miss in isolation and large enough, stacked four times, to be a different face. The model is conforming to the last thing you gave it.
Video extend has the same shape. Each extension is conditioned on the previous ending, so two extensions usually hold and five rarely look like the original. A multi-shot sequence chained last-frame to last-frame is that failure with cuts between the steps: each clip looks fine, the timeline does not.
The diagnostic is the same every time. Lay shot one and shot four next to the master still, not next to each other. Compared to each other, shot four looks like a reasonable continuation of shot three. Compared to the master, it is someone else. If your review only ever compares adjacent takes, you will sign the recast.
Two chaining jobs, two different inputs
Name the job before you pick the input.
| Job | What has to survive | Correct input | Wrong input |
|---|---|---|---|
| Continue one action past the clip ceiling | Pose, light, set, the motion already underway | Last frame of the current take (cleaned) | A new master still that resets the body |
| Start a new shot of the same character | Face, wardrobe, the person as approved | Master reference (matching angle from the sheet) | Last frame of the previous shot |
| A visible head turn or sit-to-stand inside one clip | Both identities, start and end | First-last-frame from two master-derived stills | A front master and a prompt that says "then they turn" |
| A new location, same person | Identity only; set is allowed to change | Master face and wardrobe, plus a location still | Previous clip's last frame, which also bakes in the old room |
The middle of a scene is where this gets confused. A wide at the window, then a closer face, then a profile looking down the street: that is three shots, not one motion. Each has a new camera. Each should re-anchor. Chaining the wide's last frame into the close-up gives you matching light and a face that has already started to wander, now filling the frame where it will be judged. Matching light is a location still and a lighting clause, not a dirty last frame. Reference-to-video exists so identity, set and light can be separate slots.
The re-anchor rule
Write it on the shot list, not in a style guide nobody opens.
Every new shot is generated from the master reference, plus the prompt for this shot only. The master is the still (or the small sheet of angles) you approved before clip one. It does not update when you like a take. Beautiful takes are not masters. Masters are reprintable.
- Mint the master before the sequence. A clean, well-lit still of the character. If the sequence contains profiles or backs, mint those panels too; a front master will not constrain a turn. Register them as one character so later shots point at a name rather than a folder. Reusable characters are that registration.
- Generate shot one from the master, not from a text description you will then try to honour later. Image-to-video if the shot can start on that still. Reference-to-video if the camera needs to be somewhere the still is not.
- Generate shot two from the same master, not from shot one's last frame. Change the prompt's action and camera. Do not change the identity inputs.
- Repeat until the scene is covered. If a take is unusable, regenerate that shot from the master. Do not regenerate it from the bad take, and do not feed the bad take forward so the next shot can "pick up where you left off."
- Only chain last-frame when the audience is meant to see one continuous move that the model's duration ceiling will not hold. Clean that handoff. Then, when the move is over and the next shot is a new setup, go back to the master.
Once the master is attached, stop re-describing the face. Re-description fights the still and is how you get a hybrid. Describe what is new: the action, the camera, the look. The I2V fallback chain makes the same point: the reference is the constant, the prompt is the delta.
From the master reference. Medium close-up,
three-quarter, eye-line off camera right.
She listens, then nods once. Static camera,
same overcast interior as the master.
WARDROBE: [the block, pasted verbatim]
No age, no eye colour, no "same woman as before." The still already is her. If you need a different angle of the head, swap the panel of the master you attach; do not ask a front still to become a profile across a chain of descendants.
How to keep motion without inheriting error
Re-anchoring does not mean every shot is a freeze of the hero portrait. It means identity always has a path back to the plate, and motion is requested in ways that do not smuggle a mutated face in as the plate.
New shot, new camera: re-anchor. This is most of a scene. Treat it like coverage.
Same camera, action too long for one clip: chain, then re-anchor at the next setup. Extract the last frame, clean it if you are actually extending, generate the extension, and when that action is done, the next setup goes back to the master. If you would not print that frame, do not chain it.
A turn, a sit, a pick-up that must be seen: first-last-frame from master-derived stills. Generate (or select) a start still and an end still from the master sheet, both approved as the same person, and interpolate. That is motion with two identity pins, not motion with one pin and a prayer.
Do not let a good take replace the master. Shot three is beautiful, the light is better, you start using its last frame as "the new hero." You have just moved the reprint plate to a descendant. Keep the master in a folder that does not receive exports from the timeline.
Assemble before you decide anything is the look. The editor's 480p preview pass is free, with a short per-user cooldown between previews, and a single charge applies to the final export. Watch the sequence against the master still, not against memory. Identity drift is a between-shots defect; it will not show up in a folder of individually charming clips.
If you are already four shots deep and the face has moved, do not chain a correction. Go back to the master, regenerate the drifted shots, and only then recut.
FAQ
Is this telling me to stop chaining clips?
No. Chain when you are extending one action past a duration ceiling, and clean the handoff frame first. Re-anchor when the camera, the beat, or the setup is new. Most of the compounding you are seeing is chaining used as a default instead of as a special case.
What if shot one was generated from text, with no master still?
Then shot one is not a master, it is a first draw. Still-frame the best front-facing, well-lit frame and promote that, not the last frame of the clip, to master before you generate shot two. If the best frame is already a little wrong, generate a dedicated still and start the sequence over. Starting over from a real plate is cheaper than finishing a recast.
Can I re-anchor some shots and chain others in the same scene?
Yes. A walk down a corridor that needs six seconds of the same move can be a two-clip chain. The close-up at the end of the corridor is a new shot from the master. The mistake is using one policy for the whole timeline.
The light no longer matches if I do not chain the last frame. What then?
Matching light is a lighting clause and, if you have it, a location still. It is not a reason to inherit a mutated face. Repeat the same time-of-day, direction and colour-temperature language in every shot prompt. If the model still swings, generate a locked-off plate of the empty set and use it as a set reference alongside the character master.