Cameras in meters, closed counts and named exceptions

Adiliada attaches images instead of using the asset library, states how many of everything exist in each clip, and places the camera in degrees and meters shot by shot. More than half of its recorded work is the title sequence.

Original project

This is a reverse-engineering of the prompt system behind Adiliada (higgsfield.studio), measured across 1,796 unique prompts. That is the complete history with no sampling: 229 folders and 6,677 generations read.

What this project is

A published 6-minute film. Its description reads: "A pitch-black sci-fi comedy about love, betrayal, and, above all, death. A guy named Adil accidentally stumbles into an impossible superpower: every time he dies, he's reborn in a new universe, as an alternate version of himself."

The recorded work is not spread like a film's, though. The title sequence took more than half of the project.

1,123 image prompts against 673 video prompts. Of the three films compared here, this is the only one where images dominate. A good part of that imagery is graphic design: treatments of the ADILIADA wordmark, iterated dozens of times.

Prestige TV series main title card. Pitch black frame. The wordmark 'ADILIADA'
in monumental Trajan-style Roman capitals carved from dark weathered stone,
floating in darkness, lit by a single dramatic raking light from the left that
reveals the chiseled depth of each letter and leaves the right side dissolving
into shadow. Fine dust particles drifting in the light beam. Epic, ancient,
serious — Homer's Iliad gravitas. Ultra photoreal 3D render, cinematic, no
other elements.

The prompts are short (median 1,116 characters), with one concept per prompt and always the same close: no other elements / nothing else in frame. When the deliverable is a logo, the prompt's job is to empty the frame.

The skeleton (video prompts)

One dialect, and it is the Cully Hill Boys family, but much more consistent than in ONEIRIC. The main blocks appear in 9 of every 10 prompts.

The last three are rare, but they are the most interesting in the project.

The techniques, axis by axis

1. Character anchoring: everything by image, almost nothing by library

This is the strongest departure from the other two films:

203 library references in the whole project, against 8,914 in ONEIRIC and 17,878 in Cully Hill. Adiliada barely uses Higgsfield's asset system. It attaches images to the prompt and describes them right there, with the highest 100% match rate of the three.

It also solves the USE FOR / IGNORE problem differently: the named exception. Instead of listing what to ignore, it describes the whole reference and then removes one single thing, saying explicitly that everything else stays:

<<<image_3>>>: THE HUNTER, 100% matches the reference — dark-blue distressed
long coat with brown leather panels, tan brocade waistcoat, ... and the BROWN
LEATHER CROSS-HARNESS STRAPS worn over the coat across his back. HIS FACE IS
BARE with nothing worn on it — no mask, no scarf, no goggles, no visor.
If the reference shows a face covering, that one thing is removed; everything
else is taken exactly as it is.

If the reference shows X, that one thing is removed; everything else is taken exactly as it is. This is surgery on the reference rather than a partition.

The same image can have two roles at once, and this is declared:

<<<image_2>>>: THE LOCATION AND THE FIRST FRAME OF THE CLIP — this exact
building: ... All architecture, materials, colour, light and texture in every
shot come from here.

2. Performance

Similar to ONEIRIC, and below Cully Hill on micro-behavior:

Zero literal dialogue. In what was scraped, Adiliada has no spoken scenes. The performance is physical: running, jumping, falling, dying. That is why actor direction sits inside ACTION TIMING and not in an intent block, and why the project's weight shifted toward physics.

3. Physics and logic: the strongest axis of the three films

PHYSICS is in 90% of prompts and it is full-body kinematics, not a style note:

<<<image_2>>>'s walk has real weight: heel contact, weight transfer, hip shift,
the long coat swaying with a slight delay behind each step. <<<image_5>>> has
real mass — the cylinder pulls his left arm straight, hanging vertically from
his grip on the valve, and the grey hose hangs beneath it in a loose coil,
swinging and settling slightly with inertia at each step. The cultists' robes
hang still. The eyepatch strap stretches, slides over the hair and releases as
it is pulled off, leaving a faint pressure line on the skin.

It names the whole chain: heel contact, weight transfer, hip, coat lag. And the cylinder's weight is not asserted. It is shown by its effect: it stretches his arm.

The project's signature: the closed count. 87% of prompts declare how many of each thing exist, and what each one does for the whole clip:

Exactly twelve people exist in the whole clip: <<<image_2>>>, <<<image_3>>>, and
ten <<<image_4>>>; no other figures appear inside the room. Segment 1 shows only
<<<image_3>>> in frame. The cultists' hands stay empty for the entire clip.
<<<image_5>>> stays in <<<image_2>>>'s left hand in Segments 2 and 3, always
gripped by its top valve assembly, and never changes hands; the grey hose stays
connected between the valve and the handpiece, hanging in a loose coil, the
entire clip. The eyepatch stays on his eye for the...

This attacks a typical failure of video models head on: multiplying people, teleporting objects from one hand to the other, props appearing and disappearing. The error is not prohibited. The inventory and its permanence are declared.

The CONTINUITY block does the same between clips. It starts with Direct continuation: and re-describes the inherited state of the world (who is dead, which chair is rotated 180°, where the cylinder ended up).

4. Framing and camera: the metric system

Here Adiliada is by far the most precise of the three:

The OPTICS block does not describe a look. It places the camera in space, shot by shot, in units.

LENS LOCK SHOT 1 = 65° diagonal field of view, camera about 4 meters above the
floor behind the villain, tilted down about 65 degrees — locked overhead view.
LENS LOCK SHOT 2 = 40° diagonal field of view, camera about 1.2 meters to the
DEAD MAN'S RIGHT, OFF the chair–monitor axis, slightly behind his shoulder line,
at the height of the chair's top edge, looking DIAGONALLY across the chair,
tilted slightly up, framing ONLY the crown of hair off-center in the lower-left;
then ONE continuous handheld RISE (5.0–6.4s)...

Five quantities per shot: field of view, height, lateral distance, reference axis, tilt. It is a camera sheet written in prose. And LENS LOCK per segment, no drift shuts the door on the automatic zoom that models add.

Camera behavior is directed by contrast between shots, with intensity declared relative to the rest of the scene:

Shot 1: locked-off, DEAD STATIC — not a pixel of camera movement, the horizon
LEVEL; all handheld energy is reserved for shot 2. Shot 2: REALISTIC FOOTAGE,
SHAKY HANDHELD SHOT, DIAGONAL DRIFT, RACK FOCUS. The handheld shake is VIOLENT,
THE STRONGEST IN THE SCENE — very large amplitude, high frequency, aggressively
visible in every single frame, the operator's hands never calming for a moment,
with a slow subtle diagonal drift — the horizon clearly CANTED from the first
frame, a dutch tilt of about 12 degrees...

all handheld energy is reserved for shot 2: shot 1 is defined by what it yields to shot 2. And the shake is more than "handheld". It is amplitude, frequency, drift and degrees of tilt.

73% of prompts are handheld, and 33% declare at least one locked-off shot. This is a handheld film, the reverse of ONEIRIC (stabilized anamorphic).

Two rare blocks solve problems that nobody else attacks.

SCALE ANCHORS fixes the real size of things and the camera distance, so the model reads the scale of the world:

The crab is about 15 cm across and the camera sits about 30 cm from it at ground
level, so the grey ground reads as fine gravel and small stones, not boulders.
The asteroids in segment 2 are continent-sized mountains of rock many kilometres
across.

The so ... reads as X, not Y is the key. It does not state the measurement for its own sake. It states which misreading it prevents.

HOW THE START IMAGE IS USED gives the reference image a life cycle:

The clip opens EXACTLY on the start image — the same place, the same framing,
the same composition, frame for frame. Segment 1 is that photograph coming to
life. After the hard cut the start image STOPS being a composition to copy.
From there carry over only its ATMOSPHERE: the colour palette, the direction and
quality of the light, and the character and texture of the rock. Camera position
and framing in segment 2 are free and new.

Before the cut the image controls the composition. After the cut it controls only the atmosphere. Of the three projects, this is the most elegant solution to the problem of a visual reference freezing the whole clip.

5. Texture: realism is asserted, not named

The word photoreal hardly appears, though the prompts do aim for it. They ask for it with an ontological statement at the start of SCENE CONTEXT:

THIS IS A PHOTOGRAPH OF A REAL PLACE. Every square inch of every frame is a real
physical surface recorded through a real lens: deep blue sky with a true
atmospheric gradient, real brick, plaster, timber, slate, glass, cloth, forged
steel, feathers and skin, all with photographic texture, depth and grain. Every
edge in the picture is the edge between two real surfaces, made by light.
Nothing is laid over the image and nothing floats in the air that is not a
physical thing.

Instead of asking for a style ("photoreal"), it declares what kind of object the image is. Every edge in the picture is the edge between two real surfaces, made by light goes straight at a model artifact: drawn edges, halos, things floating. It is rare in the corpus (7% of video prompts), but it is the most ambitious statement of the three films.

And light is prohibited before it is described:

Across the entire film, characters are lit ONLY by natural, soft, diffuse light
that exists in the scene — never any artificial, added or unnatural light source
of any kind (no studio lamp, key light, softbox, rim light, colored spotlight or
beam on faces).

The five moves that define Adiliada

1. Everything by attached image. 62% use <<<image_N>>>, 12% use the asset library. The reference is described inside the prompt, with 100% match in 91%.
2. The named exception. If the reference shows X, that one thing is removed; everything else is taken exactly as it is: surgery on the reference instead of a USE FOR / IGNORE partition.
3. The closed count. 87% declare how many of each thing exist and what each one does for the whole clip. This works against multiplied people and teleporting props.
4. The camera in units. LENS LOCK with degrees of field of view, meters of height, meters of offset, reference axis and degrees of tilt, shot by shot.
5. Realism as an ontological statement. THIS IS A PHOTOGRAPH OF A REAL PLACE, instead of the photoreal label.

Limitations

• The scrape is complete (229 folders, all unique prompts), but the work mix is skewed: 55% of the prompts come from the title sequence. The percentages describe that mix, from a 6-minute film and not a feature.
• SCALE ANCHORS and HOW THE START IMAGE IS USED appear in 2% and 1% of prompts. They are highlighted here for what they solve, not for their frequency. They look like local inventions for a single scene, not a convention of the project.
• The API has no ground truth for success (0 favorites, 0 published markers, no parent_id). "What worked" is inferred from the fact that the last version of each shot is the one that stayed.