Why consistency is structurally hard
Generation models do not hold a persistent identity in memory. Each clip is produced largely independently, conditioned on whatever you give it. Identity is reconstructed every time, and reconstruction drifts.
- There is no persistent character state between generations unless the model explicitly supports reference conditioning.
- Small prompt differences — a synonym, a reordered clause — shift the reconstruction.
- Motion, lighting and camera angle each add degrees of freedom in which the face can drift.
The mental model that helps: you are not retrieving a character, you are re-describing one from scratch every shot. Everything that makes the description more deterministic helps.
Techniques ranked by impact
| Technique | Impact | Effort |
|---|---|---|
| One fixed reference image for every shot | High | Low |
| Identical character wording, copied verbatim | High | Low |
| One action and one camera move per shot | Medium-high | Low |
| Locking approved key frames | Medium | Medium |
| Post-hoc correction and selective regeneration | Medium | High |
| Negative prompts and exclusion lists | Low | Low |
| Long elaborate character descriptions | Low to negative | High |
The top two cost almost nothing and do most of the work. Most people skip them and spend their effort on the bottom three.
Building a usable reference image
The reference image is the highest-leverage artefact in the whole workflow, and most references are badly made.
- Front-facing, neutral pose, even lighting. A three-quarter action shot makes a poor reference because the model inherits the angle and expression.
- Plain background. Background detail in a reference leaks into generated scenes.
- Consistent colours and no strong stylistic filter. Heavy grading in the reference propagates into output.
- High resolution and sharp. Reference quality sets a ceiling on output quality.
- One canonical image, used everywhere. Multiple references that differ slightly reintroduce the drift you are trying to remove.
Use exactly one: tempting as it is to supply several references for robustness, differing references are a common cause of drift. Pick the best one and commit.
Why identical wording matters
Paraphrasing your character description between shots is the most common self-inflicted inconsistency. The model reads wording differences as meaning differences.
- Store the character description as a fixed block and paste it unchanged into every prompt.
- Never swap synonyms between shots — “auburn hair” in one and “reddish-brown hair” in another are different inputs.
- Keep the description order stable as well as the words.
- Put the character block at the same position in every prompt so later clauses do not change how it is weighted.
Shot discipline
Every additional thing happening in a shot is another opportunity for the character to drift.
- One action per shot. “Walking while turning and smiling” is three motions competing for the same frames.
- One camera move per shot. Static or single-axis moves hold identity far better than complex moves.
- Shorter clips hold better. Four to six seconds drift noticeably less than ten.
- Keep lighting consistent between shots in the same scene. Relighting a character is functionally asking for a new reconstruction.
Locking key frames
Rather than hoping for consistency across a long generation, generate short beats, approve the frames that work, and build from them.
-
Generate the first beat and pick the best frame
Choose the frame where the character looks most like your reference.
-
Use that frame as the input for the next shot
Image-to-video from an approved frame inherits far more identity than a fresh text generation.
-
Repeat, chaining approved frames
Each link in the chain starts from something you have already approved.
-
Accept that you will re-run links
When a beat fails, regenerate that beat only. Chaining means failures are local rather than global.
Post-hoc correction
Sometimes fixing in post is cheaper than regenerating.
| Problem | Fix in post? | Better to regenerate? |
|---|---|---|
| Slight colour drift between shots | Yes — grade to match | No |
| Face subtly different in one shot | Sometimes — if the shot is short and cuttable | If the shot is held on screen |
| Outfit changed | Rarely worth fixing | Yes |
| Age or apparent ethnicity shifted | No | Yes, immediately |
| Hair length changed | No | Yes |
Know the non-negotiables: changes to identity attributes — apparent age, ethnicity, body type — are not fixable in post and read as carelessness to an audience. Regenerate those shots without exception.