What this guide helps you do
Improve continuity across AI-generated film scenes
Key takeaways
- Separate identity from appearance and state.
- Accepted frames outrank stale candidates.
- Story order matters more than generation completion order.
Film evidence
Track identity, setting, objects, and emotion separately
The robot cast needs stable identity across wide shots and close details, while other facets keep changing: corridor position, cleaning task, gaze, grouping, light, and relationship to the child's drawing. A single reference image cannot encode all of those conditions. The next shot must inherit accepted story state, not merely repeat an appearance prompt.
Generation order also cannot define continuity. A later emotional close-up might finish before an earlier routine shot, but its state must not leak backward. Canonical frames and accepted observations should update context in story order; rejected candidates, stale retries, and unreviewed outputs remain possibilities rather than production truth.
Forgotten Memories · 01:00 · Character short. This is a finished first-party Onira production, not customer proof or archive footage.
View the full film and production notesWhat to watch for
- Each robot stays recognizable as composition changes
- The drawing's location and narrative role progress causally
- Opening shots do not inherit the ending's emotional state
Section 1
Continuity has multiple facets
A character can preserve identity while changing wardrobe, injury, age, mood, or location. A setting can remain the same place while weather, damage, crowd state, and time of day evolve. A prop can move between people or change condition.
Represent those facets separately. A single reference image or repeated prompt can help appearance but cannot describe every causal state the next shot must inherit.
- +Identity and appearance
- +Wardrobe and performance
- +Setting, object, and causal state
Section 2
Accepted evidence becomes production truth
Generated candidates are possibilities. Once the editor accepts a still or motion take, observations from that result should update continuity context for later shots. Stale retries and rejected candidates must not overwrite it.
Use story order, not parallel completion order. A future scene that finishes early cannot leak a later costume, broken object, or setting condition into an earlier moment.
- +Accepted observations
- +Stable frame and take identity
- +Causal ordering
Section 3
Review continuity at boundaries
Inspect entries and exits: position, direction, gaze, carried objects, visible damage, light, weather, and final action. A transition shot may deliberately establish a new boundary; a continue-action shot should consume the previous accepted boundary at its first frame.
Continuity systems reduce drift but cannot guarantee perfect generations. Review the assembled sequence and use bounded regeneration when a mismatch breaks comprehension or character recognition.
When a replacement is accepted, record which prior frame it supersedes and which later shots consume its state. A visually better take can still be the wrong repair if it changes wardrobe, object ownership, geography, or action in ways the following sequence cannot inherit.
- +Start and final frame boundaries
- +Screen direction and object ownership
- +Human sequence review
Working standard
Publication checklist
- 01Characters have separate identity and state records.
- 02Settings and props carry evolving state.
- 03Only accepted media updates continuity.
- 04Story order controls context.
- 05Shot boundaries are reviewed.
- 06Critical drift receives bounded repair.
Primary references
Sources and further reading
Policy and model capabilities change. These sources were reviewed on July 13, 2026; open the current official page before making a production or publication decision.
Related production guides
Questions
Can the same prompt guarantee the same character?
No. Repeated prompts and references can improve consistency, but identity, appearance, wardrobe, performance, and state still need separate context and review.
What is an accepted frame?
It is a specific production frame or take selected as canonical evidence. Later continuity should use it rather than rejected or stale candidates.
Can AI film continuity be perfect?
No general workflow should promise that. Systems can reduce and detect drift, while human review and selective regeneration remain necessary.