Reference-first video is replacing the one-prompt lottery.
The important shift in AI filmmaking is not longer prompts. Production tools are becoming more stateful: creators can increasingly reuse characters, scenes and objects, control camera behavior, extend scenes and connect selected frames.
Google documents Flow features including camera controls, scene building, asset management and reusable “ingredients.” Veo documentation highlights reference ingredients, first-and-last-frame transitions, scene extension, object insertion and native audio. Read the official Flow introduction and Veo overview.
OpenAI’s Sora API guide describes creating, monitoring, downloading and remixing video jobs. That job structure is a useful signal: professional creation needs versions, status and reproducible inputs. See the official Sora video guide.
What changes for creators
- Approve neutral character and location references.
- Record identity and continuity invariants.
- Generate story keyframes before motion.
- Animate one primary action and camera intention per shot.
- Use extensions and first/last frames for transitions.
- Record every accepted version and its inputs.
What still fails
Reference support does not guarantee perfect continuity. Clothing can mutate, faces may drift under extreme angles and audio can miss editorial intention. Build coverage: reactions, inserts, environment plates and sound bridges. Test at lower cost before generating hero shots.
A useful production test
Use one character and location to create a wide entrance, medium action and close reaction. Score identity, wardrobe, geography, screen direction, lighting and sound continuity. The model producing the prettiest isolated clip may not produce the most editable sequence.