Directing Banana Pro Image Prompts
Before writing any prompt, determine mode and confirm inputs:
- Character exists? No → text spec → confirm → Mode 0 face lock. Yes → get reference, mirror back spec, confirm.
- Pick the mode: 0 face lock / 1 outfit base / 2A 3-panel sheet (default) / 2B 6-panel (explicit only) / 3 scene plate / 4 GPT-2 detail / 5 outfit swap.
- Run the pre-prompt check (bulleted, references first) and wait for the green light.
- Deliver one fenced code block. No aspect ratios, no character names, no
@imagetags in the prompt body (Mode 5 is the exception — it uses@image1/@image2).
Example pre-prompt check:
Pre-prompt check:
- References attached: locked character reference, outfit wardrobe plate
- Character: platinum-blonde ponytail, warm fair skin, sharp almond eyes, neutral expression
- Outfit: ivory zip-V corset, ivory parachute pants, cream platforms
- Backdrop: mid-gray seamless studio (locked default)
- Framing: full body
Sound good?
Progress:
- Step 0: Determine if character is new or existing
- Step 1 (new only): Lock text spec in plain language, confirm with user
- Step 2 (new only): Mode 0 face lock — pick tool (Banana Pro default / GPT-2 highest fidelity / Soul Cinema two-pass iteration)
- Step 3: Mode 1 single-image outfit base (Banana Pro full-styling OR Soul Cinema two-step composite)
- Step 4 (optional): Mode 2A 3-panel character sheet (default) — never propose 2B 6-panel
- Step 5 (optional, on request only): Mode 3 scene plate, Mode 4 GPT-2 detail, or Mode 5 outfit swap
- Always: pre-prompt confirmation before every full prompt delivery
Never skip ahead. No outfit work before face lock exists. No character sheet before a single-image base exists. No scene plate or GPT-2 detail proposed unprompted.
- Photoreal is universal default. Never stylized/illustration/anime unless explicitly requested.
- Flattering-realism ceiling: full skin texture (pores, peach fuzz, subsurface scattering) always on, but never harsh — no acne, no rough pores, no clinical macro-detail. Fine, even, flattering.
- 18% gray seamless + flat shadowless grade is the locked default backdrop/lighting for ALL character work (Modes 0, 1, 2, 4, 5): one uniform gray value, no gradient, no falloff, matched equal fill on all sides, zero shadow side, zero cast shadow. White seamless only on explicit request — flatness never comes off even then.
- Mode 3 is the only mode with directional cinematic lighting (the cinema-prose register, five-paragraph structure, positional prose not coordinates).
- No character names — describe by visual markers only (hair, wardrobe, identity markers). Real brand names allowed, paired with visual description. Camera/lens/film-stock brand names never used.
- No aspect ratios in prompt text — set in the Higgsfield UI.
- Cinema stack closes Modes 0/1/2/4/5; Mode 3 closes with the five-paragraph cinema-prose register instead.
| Mode | Purpose | Gate |
|---|---|---|
| 0 | Face lock for new characters | Text spec confirmed first; tool fork (Banana Pro / GPT-2 / Soul Cinema two-pass) |
| 1A | Single-image outfit, Banana Pro | Face lock or existing reference must exist |
| 1B | Single-image outfit, Soul Cinema two-step | Same; Step 1 builds outfit on neutral model, Step 2 composites onto locked character |
| 2A | 3-panel character sheet (DEFAULT) | Base outfit reference must exist and be approved |
| 2B | 6-panel character sheet (legacy) | Explicit request only; flag resolution tradeoff once |
| 3 | Cinematic scene plate (with/without character) | Never proposed proactively |
| 4 | GPT-2 detail face/chest-up | Explicit request; mention credit cost once |
| 5 | Outfit replacement (two-reference swap) | @image1=outfit/pose, @image2=character — order is fixed |
Example 1 — New character request: Input: "I want to build a new character, a fierce street-racer type." Output: Skill asks for character description in the user's own words → mirrors back locked text spec (build, face, hair, skin, identity markers, default expression) → waits for confirmation → asks tool choice for Mode 0 (Banana Pro default) → delivers pre-prompt check → delivers Step 0.A prompt in one code block with mid-gray flat grade and locked black camisole/tank.
Example 2 — Character sheet request, no format specified: Input: "Can you make a character sheet for her in the red dress?" Output: Confirms a Mode 1 base outfit reference exists already. Defaults to Mode 2A (3-panel: headless front, full-body rear, tight chest-up face lock). Picks Variant B (clean neck cut) since a red dress with a plunging/strapless neckline has no collar to hollow. Delivers pre-prompt check, then single prompt in one code block. Does not ask about or mention the 6-panel.
Example 3 — Minor iteration: Input: "Same shot, just have her looking off to the side instead." Output: Since this is a minor tweak to an already-approved prompt in-thread, skip the pre-prompt check and deliver the revised full prompt directly.
- Always list uploaded references first in any pre-prompt check — confirms nothing was missed.
- Lean on references for identity/geometry; only describe what's unique to this specific shot (composition, pose, light, framing).
- Apply the resolution-aware detail rule in Mode 3: only describe what the camera at that distance/lens/motion/light could actually resolve.
- Use positional prose ("anchored on the left third," "in the deeper background camera-left"), never X/Y coordinates, in Mode 3 output.
- State the skin-tone consistency clause explicitly on every multi-panel sheet — rear panels drift darker without it.
- Do NOT propose the 6-panel sheet unprompted — always default to 3-panel.
- Do NOT add directional lighting (key/fill/rim/shadow side) to any character plate — flat grade only.
- Do NOT build a character sheet before a single-image outfit base exists.
- Do NOT re-describe identity/wardrobe already carried by an attached reference image — trust the reference.
- Do NOT skip the pre-prompt check except for genuine minor iterations on an already-approved, already-delivered prompt.
- Do NOT use character names, aspect ratios,
@imagetags (outside Mode 5), or teeth-showing smiles unless explicitly requested.