2025 — Current Research
AI Cinema · Directed Production · Continuity · Precision
Source Material
Guillaume Tell · Legend c.1307
A focused research effort to validate controlled AI cinematic production through one of the most technically demanding shot sequences in the legend — the apple split. Four continuous shots. Complete character, weapon and environmental continuity across generation sessions. A model for expanding AI-native filmmaking to larger projects. Not missing the shot is both the legend's stakes and the methodology's standard.
4
Continuous
directed shots
5
Continuity
variables held
1
Saturated colour
in a grey world
Key Frame · The Apple Split · Extreme Slow Motion · Shot 04 of 04
Shot One · 4–5s
The Eye
ECU · Absolute stillness
Tell's grey-green eye along the crossbow sight. The apple — the only red — a pinpoint in the far background. The decision before the release.
Shot Two · 3–4s
The Bolt
CU · Slow tilt
Crossbow at full draw. Camera tilts from bolt to eye along the stock. The sinew tremor of held tension. Release at end of tilt.
Shot Three · 8–10s
The Flight
Follow cam · Inevitable pace
Camera 25cm behind the bolt. One hundred paces. The apple grows from pinpoint to filling the frame. Walter below it — open, unafraid.
Shot Four · 4–6s
The Split
ECU · Extreme slow motion
The apple opens — two halves rotating outward. Seeds visible one frame. They fall to grey cobblestone. Two small sounds. The world exhales.
To produce four consecutive AI-generated shots that hold complete continuity — character, weapon, environment, palette and emotional register — across independent generation sessions, reading on screen as a single unbroken camera movement.
Individual AI-generated clips can be visually extraordinary. What they cannot do without directed infrastructure is hold continuity across shots. A character's face changes between sessions. The weapon is different. The light shifts. The world palette drifts. What looks like four scenes from the same film is actually four disconnected images that happen to share a subject. The apple split sequence was chosen specifically because it demands every form of continuity simultaneously — and because failure is immediately visible.
The sequence also introduces a specific emotional challenge: the shot must feel inevitable, not spectacular. The temptation in AI generation is to reach for spectacle — dramatic light, saturated atmosphere, kinetic energy. The apple split demands the opposite. Maximum restraint. Maximum precision. One red object in a grey world. Discipline over display, at every generation decision.
Five Continuity Variables — Held Across All Four Shots
The Follow Shot
Shot 03 — the bolt in flight for 8–10 seconds — is the sequence's hardest generation problem. Camera locked 25cm behind a moving projectile, the world rushing past in correct motion blur, the apple growing from 1% of frame to 90% of frame. A continuous camera behaviour that no single generation prompt can guarantee.
The Child's Face
Walter — Tell's son — appears in shots 03 and 04. His face must carry one specific quality: not fear, not performance, not drama — the absolute calm of a child who trusts completely. Generating that distinction between performed calm and genuine stillness is one of the subtlest direction challenges in the sequence.
The Physics of the Split
The apple must open, not explode. Two halves rotating outward with botanical precision — not shattered fragments, not dramatic burst. The seeds visible for one frame. This is an extremely specific physics event that generative systems default to spectacularising. The direction must work against that instinct.
The Final Image
Shot 04 ends on two red half-circles on grey cobblestone — held for one full second. This is the image the entire sequence is built toward. Its emotional weight depends entirely on what precedes it — and on the restraint not to undercut it with anything decorative in the final frames.
Four tools in strict sequence — intellectual infrastructure first, visual architecture second, video generation third, assembly and grade in a professional editorial environment.
↳ The Pipeline Logic
This pipeline has a precise division of labour. Claude Project completes all direction before any tool generates a single frame. Luma AI acts as the visual orchestration layer — establishing the flow and generating keyframes (first and last frame per shot) using its Nano, Banana and Pro model nodes. Those keyframes feed directly into Seedance 2.0, which generates the video between them — the motion, the flight, the split. DaVinci Resolve assembles and grades. Each tool operates at its ceiling rather than outside it.
Claude Project World Context Document
A persistent world brief written to anchor every tool in the pipeline — palette rules, light direction, camera behaviour philosophy, scale logic, texture register, style references (Friedrich, Ghibli, Tarkovsky, The Revenant), and explicit negative prompts. Injected with every Luma session and every Seedance generation run. The core rule is stated and restated: the world is desaturated grey. The apple is the only red. Nothing else.
Claude Project Character & Environment Bibles
Full creative documents for every visual element that must hold continuity across generation sessions — Tell's physical attributes and emotional register, Walter's specific expression quality (not fear, not performance — the calm of absolute trust), the Altdorf square's spatial character and colour palette, the crossbow's material description. Each is written as a generation-ready brief and stored as a reusable component. Character identity does not drift between Luma and Seedance sessions because the brief is written to prevent it.
Claude Project Shot-by-Shot Direction Documents
Each of the four shots receives a full breakdown: shot type, camera motion, subject in focus, background depth, emotional register, duration, and both a Luma keyframe brief and a Seedance motion brief. The follow shot (S03) receives the most detailed direction — motion blur behaviour, the approach speed, the apple's scale progression from 1% of frame to 90%, and the specific quality of the child's face as the bolt arrives. Negative prompts are as detailed as positive prompts.
Luma AI Visual Flow & Keyframe Generation
Luma AI operates as the visual orchestration layer — establishing the overall cinematic flow and generating the anchor keyframes for each shot: the first frame and last frame that define what the shot begins and ends on. Luma's node-based model stack — Nano for rough exploration, Banana for motion planning, Pro for final keyframe fidelity — allows progressive refinement from concept to production-grade reference frame. These keyframes are not suggestions. They are the visual contract each shot is held to.
Seedance 2.0 Video Generation Between Keyframes
The Luma-generated keyframes feed directly into Seedance 2.0, which generates the video between them — all motion, all physics, all camera behaviour. Seedance receives the first frame, the last frame, the world context document, and the shot direction brief. It generates the bolt's flight, the apple's split, the slow-motion fall. Multiple variants are generated per shot — selected against continuity, emotional register, and camera behaviour. The follow shot (S03) requires the most iterations: the camera-behind-the-bolt movement is the pipeline's hardest generation request.
DaVinci Resolve Assembly, Colour Grade & Final Cut
Selected clips are brought into DaVinci Resolve for timeline assembly, colour grade, and final cut. The editorial work establishes the sequence's rhythmic architecture — the held stillness before the release, the silence of the flight, the slow-motion of the split, the full second on the fallen halves. The colour grade enforces the visual rule at the technical level: a selective desaturation node pulls every colour toward grey except a precisely defined red range. The apple's red is the only value that survives the grade. What every upstream tool works toward, the grade delivers. The cut is the argument.
A proof of concept — four AI-generated shots holding complete continuity and emotional coherence, assembled into a 24-second sequence that demonstrates directed AI cinematic production as a viable methodology for larger projects.
4
Shots with full
continuity held
24s
Sequence
duration
1
Red object
in a grey world
∞
Pipeline scalable
to full film
The core finding of this research is precisely located: visual continuity across AI generation sessions is a direction problem, not a tool problem. When the character, weapon, environment, palette and emotional register are fully specified in writing before generation begins — and when Luma AI's keyframes give Seedance 2.0 a precise visual contract to interpolate within — the resulting clips hold continuity with a fidelity that produces a coherent sequence. When the brief is partial, continuity breaks. No amount of regeneration fixes under-direction.
The keyframe-to-video handoff between Luma AI and Seedance 2.0 is the pipeline's most technically productive insight. Luma establishes the what — the first and last frame of each shot. Seedance generates the how — the motion between them. Separating visual architecture from motion generation gives the director two precise intervention points rather than one opaque prompt. The apple's split, the bolt's flight, the child turning — these are Seedance events, constrained by Luma's keyframe contract and directed by Claude's brief.
The DaVinci Resolve colour grade is not a finishing step — it is the pipeline's final creative argument. The apple split is the proof. The legend is the project.
"Not missing the shot is not a metaphor. It is the specification. The legend set the standard. The methodology meets it."
Vertical Labs · Culture & Entertainment · Not Missing the Shot · 2025