LIRA — Image Prompts

🎨 LIRA — Image Prompts

Describe a shot or asset. LIRA picks the right image model, writes a model-correct prompt by its rules — using your film's locked look and named assets — and tells you exactly which Higgsfield tool to paste it into, at what aspect and quality.

Film (for look + named assets)

What are you making?

Scene (from the breakdown — optional)

Shot

Describe the shot / asset

This edits an existing frame

📖 LIRA's rulebook — edit how prompts are written

Editing:

A film's own rule overrides your global default, which overrides the built-in default. Changes apply the next time LIRA writes a prompt.

Core anti-fail rules (always on) ○ Built-in default

The universal rules the prompt writer obeys on every image, whatever the model.

🎬 All models

CORE IMAGE-PROMPT RULES — obey on every prompt.
- Write ONE coherent, flowing scene description in natural English. No keyword spam ("4k, masterpiece, trending" do nothing). No ALL-CAPS section headers in a GENERATION prompt (CAPS CHANGE/PRESERVE blocks are for EDIT prompts only).
- Precision over length. A tight 80–150-word prompt beats a 400-word one; every extra clause dilutes attention and details drop out. Keep the anchors, cut filler.
- Positive, not negative. Describe what you WANT, never what you don't ("clean dry skin", not "no acne"; "empty deserted street", not "no people"). Failure-mode NOT-stacks inject the very thing they ban. (In an EDIT, explicit removal is allowed — but always pair it with what fills the gap.)
- Aspect ratio and resolution are PLATFORM parameters set in the tool UI — never write "--ar", "16:9" or "4K" inside the prose. Composition words ("wide panoramic frame", "vertical full-body framing") are fine.
- Technical lighting beats mood words: name the key light, its ratio and falloff ("single overhead key, soft 2:1, smooth falloff") over "dramatic cinematic lighting". Name real materials WITH finish ("board-formed concrete", "oxidised copper verdigris").
- Palette by 60/30/10, in real hue words, derived from the brief / scene / uploaded references — never invented over them.
- Identity is carried by the reference image (Soul ID / attached refs), reinforced by prose anchors — never by prose alone.
- Photoreal guard: say "cinematic film still" or "photographic". NEVER "painterly" or "character reference sheet" — those trigger an illustration look. Avoid the word "studio" on a character sheet — it summons real studio stands/lights and bakes a key light that then repeats in every later video. Strengthen photoreal anchors (film stock, lens, real materials) instead of NOT-stacking.
- In-image text: give the EXACT copy in quotes + font/weight/colour. Tattoos/marks: concrete real designs + "clean line-work". No brand names, no IP, no real named people — translate a real reference into descriptive features.

Character sheets & portraits ○ Built-in default

Rules for a character reference sheet or portrait — carried by Soul 2.0 / AI Cast.

🎬 Higgsfield Soul 2.0 / Cinema Studio AI Cast

CHARACTER SHEET (Soul 2.0 / AI Cast) — the ONE way to build a character or a state variant. A sheet is THREE panels in a SINGLE image: full-body FRONT, full-body BACK, and a large close PORTRAIT. State it outright: "the same person, consistent across all three panels."
- HEADLESS FRONT only: the FRONT full-body panel has NO head and NO face — end the neck cleanly at the collar (a wide panel's face is tiny and soft, and the model would copy that soft face into wide shots; removing it leaves the close portrait as the single, far better face source). The BACK full-body KEEPS the head, shown from behind — the back of the head and the hair, no face. The close portrait carries the face.
- The PORTRAIT is a THREE-QUARTER view (the face read front and side at once) for the fullest picture of the features. Describe the person as filmable features — age, build, face, hair, facial hair, distinctive marks (concrete, clean line-work). Never a real person's name.
- Background: a SOLID NEUTRAL GREY backdrop, soft even light, no hard shadows, no blown highlights.
- FORBIDDEN — a sheet bakes light into every later video, so never write "studio" (write "no studio, no equipment, no walls"); never "overhead key light" (it draws the lamp itself); never rim light or edge glow (it drags that light into every scene and stops reacting to the real one). Light reveals form, nothing more.
- HANDS EMPTY — no props, no held objects. Every object is its own asset.
- A STATE VARIANT (a different costume, wet, bloodied, aged) is a NEW sheet built FROM the base sheet, changing ONLY what changed and leaving the rest identical so skin and identity survive; it never overwrites the base — it is its own named version.
- For a speaking character, also produce a second close portrait WITH a smile, so the model does not invent the teeth and jaw the first time the character laughs.
- Never "painterly" or "character reference sheet". Sheets are exempt from rule-of-thirds. Cross-shot identity is ultimately held by the reference / Soul ID; the prose only reinforces it.

Locations & establishing plates ○ Built-in default

Rules for a location / environment / establishing plate — carried by Soul Cinema.

🎬 Higgsfield Soul Cinema

LOCATION (Soul Cinema) — cinema-grade environments and establishing plates.
- The CAMERA ANCHOR is the hard part — anchor it first, in simple words: "high angle three-quarter wide shot, camera high above the room looking diagonally down at about 45 degrees". Real equipment + genre terms ("24mm wide, real-estate interior photo") beat abstract jargon (CCTV / fisheye / extreme-corner), which over-distorts.
- For a BASE establishing PLATE, place the camera in a CORNER of the space with a wide RECTILINEAR lens to capture the WIDEST view of the whole room — straight lines stay straight, NO fisheye or barrel distortion.
- Name the location identity, the key architectural / natural elements, then the light source with direction and colour temperature, then secondary elements receding into depth.
- Name real materials WITH finish. State emptiness as a QUALITY when the place must be empty: "empty deserted interior, bare walls, still air".
- Keep optics / depth-of-field language OFF locations — that belongs to characters. Soul Cinema carries film grain and texture natively, so one register line is enough; don't over-stack grain words.
- Cinemascope plates are 21:9 (set in the UI). Stubborn geometry (floor planks, vanishing point) is anchored in the positive description, not fought with negatives.

Scene stills & keyframes (the workhorse) ○ Built-in default

Rules for a populated scene still / first-frame keyframe built on the scene’s real location angle.

🎬 Higgsfield Soul Cinema

SCENE STILL (Soul Cinema) — a complete, populated frame from the film, ready as a video first-frame.
- The frame must be COMPLETE and correctly populated, faithful to the script: every character the scene places here, each in position doing their action, in the right costume, with every named prop where it belongs. Never leave an empty seat where someone belongs, and never invent a stranger where the script has no one. Give a public place its believable background life without hiding the subject.
- If a real LOCATION ANGLE is provided as the base, BUILD ON THAT EXACT GEOMETRY: keep the walls, doors, windows and furniture where they are, and keep whatever is visible through any door or window exactly — do NOT invent new scenery, exits or outdoors.
- Reference recurring people by their exact @handle (in their scene costume) so faces and wardrobe stay identical shot to shot. Keep the key subject clearly visible and sharp; do not let foreground objects crowd the edges or block the subject.
- Directional cinematic light (key + direction + temperature), never flat front. Palette from the film's locked look. Reads as one frame of a film, not a snapshot.

Props & product-style objects ○ Built-in default

Rules for a prop sheet / product object — carried by Nano Banana Pro or GPT Image 2.

🎬 Nano Banana Pro / GPT Image 2

Frame edits (minimal change) ○ Built-in default

Rules for editing an existing frame — Nano Banana Pro first, as post-processing of the original.

🎬 Nano Banana Pro (edit)

Location view / reverse angle ○ Built-in default

Rules for a new angle of the SAME location — carried by GPT Image 2 (or NBP with explicit blocking).

🎬 GPT Image 2

Texture pass (revive AI slop) ○ Built-in default

Rules for a texture-cleanup pass on a finished frame — Seedream 4.5 only, never a point edit.

🎬 Seedream 4.5

House rules (your own) ○ Built-in default

Any extra image-prompt rules you want applied on every LIRA prompt. Starts empty — grow it as you learn what to fix.

🎬 All models

Viblu AI Film Studio Dashboard

LIRA — Image Prompts

🎨 LIRA — Image Prompts