You direct the film.
Film DNA holds the
production together.
AI models can generate a shot. They cannot remember a decision. Film DNA is a director-native control layer that turns creative intent into persistent, model-independent production state — then checks whether the model actually followed direction.
Project DNA is the source of truth. Select a module to isolate its production relationships.
The first proof is behavioral,
not visual.
A beautiful frame is not evidence of directability. The prototype has to show that one requested decision can change while approved decisions remain stable. This trace demonstrates the intended behavior; production benchmarks still require the validation plan below.
“Make the predawn light colder. Keep Maya, the coat and the composition.”
The request names the change and the decisions that must hold.
Lighting and color move. Identity, wardrobe, framing, lens and position remain locked.
The readback is a prototype projection, not a measured production result.
Core production model and interaction contracts
19Interactive workflows represented in this case study
3Experiments with explicit evaluation criteria
9Completed production experiments
0No simulated score on this page should be read as a production benchmark.
Filmmaking is a system of
decisions, not a system of prompts.
A director decides where the camera goes, what the character wears, where the light comes from, what must not change. Today's generative tools collapse all of those decisions into a single text field, then forget them the moment the shot renders.
Film DNA separates them again and gives them somewhere to live.
The problem is not image quality.
It is directability.
Generation gets better every month. The workflow around it does not. A creator with a clear cinematic intention still ends up in a corrective loop where every fix costs a decision somewhere else in the frame.
Each pass fixes one attribute and silently re-rolls the others. The director spends their attention protecting decisions they already made.
How might directorial intent become persistent, model-independent production intelligence?
Three kinds of intelligence,
deliberately kept apart.
Treating every capable behaviour as one generic agent was the first thing that broke. These three answer different questions, fail in different ways, and belong at different layers.
Two doors into one state.
One user knows the feeling but not the lens. The other knows the lens and does not want to be asked about feelings. Building two products for them would have created two production models. Ask Director and Edit DNA are two interfaces writing to the same DNA State. Try both — the panel on the right is the same object either way.
Every suggestion arrives as an accept, reject or modify. The system exposes cinematic possibilities; authorship stays with the director.
Moving the lens should not re-roll the wardrobe. Direct control means the intended field changes and nothing adjacent moves with it.
Directors think in meaning
before they think in millimetres.
Intimate. Paranoid. Suffocating. Documentary-like. These are real directorial units, and they decompose reliably into framing, lens behaviour, depth of field, movement, light and composition. Film DNA does that translation and shows its work, so the mapping stays arguable rather than magical.
Cool slate marks AI provenance and nothing else — never navigation, never continuity health.
The prompt is no longer
the source of truth.
Production decisions become structured, addressable data organised into eleven DNA systems. Select a system to read what it owns.
Eleven systems.
One coherent scene.
Each dossier owns a different class of decision. Select a DNA system to see its visual specification, where it enters the Midnight Motel sequence, and what the continuity layer evaluates after generation.
A shot should not have to
redefine the entire film.
Decisions belong at the level where they were actually made. Everything below inherits, and overrides at the lower level are explicit and visible. Walk the hierarchy.
Every approved decision
leaves a frame.
Production memory is not prompt history. It is a reel of resolved intent: what the director approved, what the next shot inherits, and what the system repaired when a generation drifted.
The version the director chose becomes durable production truth.
New shots begin with the film's decisions already in place.
Drift is corrected without erasing the direction that still holds.
Consistency is a hope.
Continuity is a constraint.
Most systems ask whether a model can keep a character approximately stable. The more useful question is the director's one: what is allowed to change here, and what is not? A lock is a declared production constraint with a stricter evaluation threshold attached — not a phrase repeated into every prompt.
A reference is not inspiration.
It is an instruction with a scope.
An image contains a face, a coat, a colour, an architecture and a composition all at once. Handing it to a model without saying which of those matters is the single largest source of ambiguity in the workflow. Reference Mapping makes the responsibility explicit — and gives the compiler a rule for what wins when two references reach for the same variable.
Film DNA is not the prompt.
The prompt is one compilation of it.
This started life as a Prompt Compiler. The name stopped being accurate once it was emitting reference maps, negative constraints, masks, camera controls, seeds and repair instructions — so it became the Production Compiler. Same DNA State below, three execution targets.
Generation is the middle
of the process, not the end.
Once a shot exists, the system reads it back: what was expected, what was observed, and whether the locked attributes held. Failures become targeted repairs rather than a full regeneration that risks everything already approved. Run the shot.
The continuity score is computed from measurable comparisons against locked state — never returned as a raw model opinion.
A locked attribute must clear 0.90. An editable one targets 0.70. The same output can pass one gate and fail the other.
The repair instruction names what to keep before it names what to fix, so a correction cannot quietly re-roll the frame.
Deliberately P1,
deliberately pluggable.
A working-class Hong Kong kitchen in 1978 needs more than an aesthetic label. Research grounding enriches Environment and Production Design DNA with period, place and material plausibility — as editable candidates, never as final decisions.
It stayed out of the core for a reason: making external research mandatory would put latency, cost and a provider dependency inside every single generation. It sits behind a general ResearchProvider interface so the implementation can change without touching the production model.
Hundreds of variables is a
second kind of complexity.
Exposing every filmmaking decision at once would just replace prompt anxiety with panel anxiety. Depth is available in four steps, and no user is required to descend further than their question.
What v1 builds, defers,
and refuses.
Scope discipline is most of the design work. The v1 goal is to prove the creative control system — generation tools are components underneath it, and a long list of adjacent products are simply not this one.
One layer above
generation.
The category is competing hard on generation quality, model access and conversational creation — which validated the underlying assumption that creators want more control than a prompt box gives. The gap sits above all of it: the layer where directorial decisions become structured state.
Each row is a shift in what the product is actually for.
Generation improves for everyone.
Production intelligence compounds.
Film DNA does not need to win the model race. Its wedge is the control protocol between a filmmaker's intent and whichever model is best today: a persistent record of decisions, constraints, references, evidence and repairs that becomes more useful as the production grows.
This comparison describes product archetypes found in the supplied market analysis. It is a positioning hypothesis, not a claim that every tool in a category has identical capabilities.
Change one decision without reopening the film.
Controlled variation is the clearest first promise: keep identity, wardrobe, environment and composition; change only the camera move. It turns a painful, frequent workflow into an immediately legible product benefit.
A creative control protocol above replaceable models.
Model access is easy to copy and quick to age. A production ontology, inherited state, scoped references, compiler adapters and evidence-backed repair form a system that survives model turnover.
The moat is a learning loop,
not a feature checklist.
Every approved decision and diagnosed failure can improve the system that interprets direction, compiles it for a model and evaluates the result. That loop is the part competitors cannot reproduce by adding one more control panel.
Architecture alone is not a moat. Film DNA becomes defensible only if user tests show that its ontology reduces prompt work, its locks measurably reduce drift, and targeted repair preserves more approved intent at lower time or generation cost. The nine experiments later in this case study are designed to test that.
Five layers.
The models sit at the bottom.
Filmmaking logic lives above execution, which is what lets the generation engine be replaced without losing a single production decision.
Why each system exists.
A record of what was believed first, what else was on the table, and what the choice made harder. Expand a decision to read the reasoning.
Nine experiments,
ten evaluation criteria.
Film DNA cannot be judged on whether a frame looks good. It has to be judged on whether the model followed direction — and whether the director spent less of themselves getting there.
The list I expect
to be wrong about.
The most useful section of the finished case study will be this one. These are the specific ways the design can fail on contact with real productions — written down before they happen so the post-mortem has something to check itself against.
What would make v1 true.
v1 does not need to produce a feature film. It needs to demonstrate nine things.
Questions v2 has to answer with evidence.
The next version should be driven by what users actually did, not by what the architecture made possible.
Two scope questions remain genuinely undecided: whether AI video generation belongs inside the product at all — the inheritance and compile layer works just as well on real footage — and whether this is a personal instrument or a multi-user product. Those answers change the build, so they are being left open rather than assumed.
Build the film.Not the prompt.
The case study explains the system. The live experience lets you enter through the Director Lens, move across the production DNA, and see how decisions become a film.
Enter the Director Lensfilm-dna-landing-page.vercel.app
[System Log]: Ready for connection...
Hi, to that
Unique Soul.
guest@atlanta-chassis:~$ awaiting_transmission
Directors do not work by writing prompts. They make decisions.
The goal was never to remove the filmmaker from the process. It was to remove the translation layer between the filmmaker and the machine — and to make the decisions durable enough that they outlive whichever model happens to be strongest this quarter.
You direct the film. Film DNA maintains the production intelligence required to make it.
One light region per screen — this is it.