How the work gets made
Brief to finished footage. Frames get locked first and only then are allowed to move, which is the whole argument: a video model asked to invent a scene will warp and drift, and the same model handed an approved frame will animate it. Each stage locks something down so the next one cannot wander, and each exists because of a specific way the work goes wrong without it.
The pipeline
Stages one to six produce a frame somebody has signed off. Stages seven to nine turn signed-off frames into footage. Nothing is allowed to move until it is right as a still, which is why the motion half is short.
- 1
Brief intake
Locks: What the thing is, before anything is made
The brief becomes a written spec: core concept, a master style block every shot inherits, palette assigned by role, world down to costume and set dressing. One source, so no frame gets re-invented later.
- 2
Reference capture
Locks: Identity, as an image rather than a description
Anything that has to stay the same across the set enters here as a reference, not as wording. Faces, products, geometry. This is the stage that decides whether the set will hold together, and it happens before the first generation.
- 3
Generation
Locks: The frames themselves
Prompts assembled from the block system against the locked references. Keepers are recorded with everything needed to reproduce the decision, so a good result is an asset rather than an accident.
- 4
Consistency check
Locks: That it is still the same subject
A gate, not a glance. The set is compared against the reference, and anything that has drifted goes back rather than forward. Catching it here costs one regeneration; catching it in the cut costs the sequence.
- 5
Retouch
Locks: Pixel-level correctness
Defects are fixed in place rather than re-rolled, because re-rolling gambles the whole frame to fix one part of it. Every correction is recorded: what was wrong, what was done, why that beat regenerating.
- 6
Dimensional and optical QA
Locks: That the image obeys physics
The stage most pipelines do not have. Scale, perspective and horizon are checked against stated rules, and lighting is audited for whether every shadow in the frame implies one source. These are pass or fail, which means a client can check them too.
- 7
Image to video
Locks: Motion, bounded by frames already approved
Approved stills move. Camera direction is written as shot grammar rather than mood, naming the move explicitly: slow push in, left glide, low-angle tilt. Where the shot allows it, motion is bounded by an approved frame at both ends so the model interpolates between two known images instead of inventing between them.
- 8
Shot correction
Locks: The take itself
Generated footage gets treated as footage. When one element in a shot is wrong and the rest is right, it is corrected video to video on the existing clip rather than regenerated, because generated takes are not reproducible. Re-rolling does not return the same shot improved, it returns a different shot.
- 9
Assembly and handoff
Locks: That somebody else can pick it up
The sequence is assembled as a deliberate edit rather than accepting a single machine output as the take. Naming, versioning and the written record travel with the files. The test of a pipeline is whether the next person can continue it without the first person in the room.
The prompt architecture
Prompts are assembled from ordered blocks rather than written as sentences. The order is fixed, so two people building a shot a week apart produce the same structure, and a shot that comes out wrong can be diagnosed to a block instead of rewritten from scratch.
01
Subject
what the thing is
02
Material
what it is made of
03
Environment
where it lives
04
Light
one direction, named and held
05
Camera
focal length, which sets spatial feel
06
Render
the closing qualifier
One light direction is chosen at the brief and held across every frame in the set. It is the cheapest rule in the system and the one that does the most work, because inconsistent light is the thing an audience notices without being able to name.
Above all of it sits one finding that governs the rest: a reference image beats a description every time. Words are reliable for wardrobe, staging and light, and unreliable for identity, because a description of a face returns a category of face. So the text carries style and the reference carries the person.
Failure modes the stages catch
A pipeline is only worth the failures it is built around. These are the recurring ones and what each actually indicates, which is the difference between fixing a shot and re-rolling until something works.
It looks illustrated, not photographic
Control strength pushed too high, or too many controls stacked against each other.
The geometry moved
No depth control, or its weight is too low to hold the structure.
Nothing changed at all
The opposite reading. Strength too low, or one control dominating the rest.
The edit bled outside the mask
Instruction-based tools regenerate wider than the area you pointed at.
Text came out garbled
Never let a model generate type. Composite real type over the plate.
What the structure is for
Generated frames are cheap, so iteration is cheap. What stays expensive is judgement: deciding what the thing looks like, catching the shot that has quietly gone off model, and knowing which stage a problem belongs to. A fault in the cut is usually a fault in the spec.
Running it as stages rather than one long session is what makes the work repeatable by someone else and reviewable by someone who was not in the room. That is the difference between a person who can generate images and a process a team can put a deadline on.