Higgsfield Camera

Lens + aperture by shot purpose

[OFFICIAL — Higgsfield shotlist-builder skill, 2026-07; re-authored from the Chinese source] — pick the glass from the shot's purpose, then say both lens and aperture in the prompt:

Two standing clauses that ride with these choices:

• Focus lock on inserts and tight close-ups: "focus plane locked on [object] from first frame to last — no focus drift, no rack focus, no autofocus hunting."
• Distortion forbid on wide/fast glass: "straight rectilinear lines, no barrel or pincushion distortion, no fisheye, no wide-angle warp." (For Seedance block prompts, FOV is stated in degrees instead of mm — ../higgsfield-seedance/SKILL.md § FOV anchors; this table's mm/aperture vocabulary serves Cinema Studio and non-Seedance surfaces, and the purpose→glass logic transfers either way.)

Shot duration by type

Durations inside a 15s multi-shot envelope — divide the envelope using these, and let an emotional arc earn its full length:

Micro-moves need exact distances

For very slow dollies, state total travel and time, or the model oversells the move: "over the full 7 seconds the camera pulls back only 10–15 centimeters — slow enough to barely register. No zoom, no sudden push." Never write zoom for a physical move — name dolly / track / push-in / pull-out.

And run the check in reverse: the stated move must be able to produce the stated end framing. "A slow push" cannot end on an extreme close-up from a wide, and a 10cm pull cannot land a full-body reveal. When move and end frame disagree, the prompt is internally impossible — the model picks one and improvises the other. Fix it on whichever side is cheaper: enlarge the move (name the bigger travel) or shrink the framing change. [EMPIRICAL — third-party director-skill evaluations 2026-08-09; re-derived]

Cinema Studio 3.0 Camera Best Practices (Business/Team Plan)

These best practices apply to Cinema Studio 3.0's generation engine, available exclusively on Business and Team plans.

The One-Move Rule

For any single shot, specify only ONE primary camera move. Do NOT stack multiple moves (e.g., dolly push + pan left + tilt up). This is the #1 cause of jitter, unwanted rotation, and failed generations.

Wrong: Camera: dolly push forward while panning left and tilting up to reveal the skyline Right: Camera: slow dolly push from medium shot to tight close-up over 8 seconds

If you need multiple camera moves, break them into separate shots using Cinema Studio 3.0's Custom multi-shot mode.

Genre-Based Camera Presets

Reliable Phrasing Library

These phrases produce consistent, predictable results:

Camera Transfer via @Video Reference

The safest way to achieve complex camera motion is to clone it from a reference:

Match the camera movement from @Video1. A dancer performs on a rooftop at sunset.

This bypasses the One-Move Rule because the model extracts camera data directly from the reference rather than interpreting text instructions.

Dual Video Reference

Action reference and camera reference can come from DIFFERENT videos. Separate them clearly:

Reference @Video1 for the character's movement and choreography.
Reference @Video2 for camera movement only.
A martial artist performs a spinning kick in a dojo.

Smart Mode (Cinema Studio 3.0)

Cinema Studio 3.0's "Smart" shot control delegates camera planning to the model. When you select Smart mode:

• The model auto-plans camera language based on the genre and scene description
• Trust it for genre-appropriate camera work — describe the feeling or genre rather than specific camera moves
• Best for: users who want professional-looking camera work without manual specification
• Override by switching to Custom multi-shot for per-scene camera control

Smart mode prompt example:

Genre: Drama. A woman sits alone at a café table, stirring her coffee absently.
She notices someone through the window and her expression shifts from sadness to surprise.

(No camera instruction needed — Smart mode will select genre-appropriate drama camera work.)

Video Reference — What It Reads, and What It Can't

The repo documents @Video reference usage in several places (Camera Transfer above, higgsfield-motion for action choreography, higgsfield-style for grade carryover, higgsfield-audio for voice timbre, higgsfield-cinema for the Motion/Camera Sheet). The capability boundary — what a video reference reads reliably, where it falls short, and what it cannot do at all — is documented once here and applies across all those use cases.

What It Reads Reliably

A @Video reference reliably carries these properties from the source clip into the generation:

• World design — the overall environment, set construction, lighting logic, color palette, atmosphere.
• Material and texture — how surfaces look (wet wood, rope, water behavior, fabric weight).
• Physics and motion style — how objects move, how water behaves, the weight and energy of action.
• Camera character — the overall feel of how the camera operates, its proximity to subjects, its movement vocabulary.
• Production quality level — high-budget vs. low-budget feel, the overall cinematic register.
• Color grade — tonal treatment, contrast, saturation level.
• Weather and environmental conditions — storm intensity, rain density, fog, light quality.

What It Reads Less Reliably

These properties carry across inconsistently. Don't depend on them when designing the prompt:

• Frame-accurate action continuation — asking the model to pick up exactly where the last frame ended is unreliable. The reference understands the energy and world of what was happening, not the precise final position of every element.
• Exact character identity — use an image reference for character identity locks, not a video reference. Video references read world and energy better than face or costume precision.
• Specific dialogue or audio — voice character can carry over but exact vocal performance is not guaranteed.

Four Reference Patterns by Purpose

Make the reference's job explicit in the prompt. Pick the pattern that matches what you want the reference to do:

The pattern works because it tells the model which axis of the reference matters and lets the rest of the prompt drive the new content.

Hard Limits

These are not "less reliable" — they are things a video reference cannot do at all:

• Cannot continue a specific action from the last frame. If the reference clip ends on a wave hitting a ship, the new clip will not automatically start with that wave still hitting. The reference supplies world context; write the new clip's opening action from scratch.
• Cannot override your prompt. The reference informs the generation — it does not replace prompt instructions. Where prompt and reference conflict, the prompt generally wins on action and the reference wins on texture and world feel.
• Cannot supply character identity alone. Always pair a video reference with an image reference if character consistency matters.

Load-Bearing Rule

Prompt wins on action, reference wins on texture and world feel. Write the prompt for what happens; let the reference carry how it looks. Pair with an image reference when character identity must hold.

Related skills

• higgsfield-seedance-vfx — Video-to-video footage transforms: preserve a real handheld/driving camera move frame-for-frame while swapping the world or adding an effect, and add a timed camera move (crash-zoom, push-in, reveal pull-back) you never actually filmed
• higgsfield-motion — Named motion presets (VFX overlays applied with camera moves)
• higgsfield-cinema — Director Panel (18 Cinema Studio camera movements)
• higgsfield-image-shots — Camera angles and implied movement for still images
• higgsfield-prompt — MCSLA formula, prompt structure
• higgsfield-style — Visual styles to pair with camera choices
• templates/ — Annotated genre-specific prompt templates demonstrating camera use