Last updated: 2026-07-13
A generation is only as good as the sentence you feed it. In 2026 the frontier models — Midjourney V7/V8, Google's Nano Banana Pro (Gemini 3 Pro Image), and Black Forest Labs' FLUX.2 — all reward the same discipline: a scene described in ordered, affirmative, camera-literate language, then locked with references, seeds, and style codes so a one-off becomes a repeatable set. This chapter gives you the slot-by-slot anatomy of a strong prompt, the exact parameters that lock consistency, per-model cheat-sheets, and the real reasons prompts get silently rejected — with copy-paste templates for each.
What matters most
- Every top model in 2026 (Nano Banana, FLUX.2, Midjourney) converged on the same skeleton: Subject + Action/State + Setting + Composition + Lighting + Style/Medium. Fill all six slots in that order — word order is load-bearing.
- Front-load the subject. FLUX.2's VLM and Nano Banana both prioritize whatever comes first, so 'A weathered fisherman...' beats 'In a stormy harbor, there is a fisherman...'
- Use affirmative phrasing, never negation. 'A well-lit room' outperforms 'a room that is not dark' on every VLM-based model. Save exclusions for the model's dedicated negative slot (Midjourney --no), not the prose.
- Trade adjectives for materials and specifics: not 'a suit jacket' but 'navy blue tweed'; not 'armor' but 'ornate elven plate etched with silver leaf'. Concrete nouns carry more signal than piled-up adjectives.
- Speak camera, not vibes. Name lens + aperture (35mm f/1.8), shot size (medium-full, center-framed), angle (low-angle), and film stock (medium-format analog, 1980s color film). These terms deterministically move depth of field and mood.
- Specify light as type + direction + quality: 'golden-hour backlight through mist, long shadows' or 'three-point softbox, even' or 'chiaroscuro, harsh high contrast'. Generic 'warm lighting' wastes a slot.
Common mistakes to avoid
- Don't stack synonyms ('beautiful, gorgeous, stunning, pretty') — models flatten them; one precise descriptor wins.
- Don't randomly list elements. Describe foreground → midground → background in order; chaotic order confuses spatial layout, especially on FLUX.
- Avoid prompt-weight syntax like '(word)++' or 'word:1.4' on FLUX/Nano Banana — unsupported; write 'with emphasis on ___' instead. Weights only work in Midjourney/SD-lineage tools.
- Don't trust seeds as a hard consistency tool — Midjourney's own docs say seeds have the least impact and 'may behave unexpectedly' across sessions/versions (only ~99% identical). Use references/style codes for real identity locking.
The short version
- Fill six ordered slots — Subject, Action, Setting, Composition, Lighting, Style — in affirmative, camera-literate prose; word order and front-loading determine what the model prioritizes.
- Trade adjective piles for concrete materials, named lenses/apertures, and specified light (type + direction + quality). 'Navy tweed, 85mm f/2, golden-hour backlight' beats 'nice fabric, cinematic'.
- Lock consistency with references and style codes, not re-description: Midjourney --sref (--sw 0–1000) + Omni Ref; Nano Banana named references + conversational editing; FLUX.2 native multi-reference. Seeds only pin noise and are unreliable across sessions.
- Always set aspect ratio and compose at native resolution, then upscale last; match ratio to platform (9:16, 4:5, 16:9). Quote exact text strings and name the typeface for reliable typography.
- Pick the tool for the job: Midjourney for stylized series, Nano Banana for edits/text/fusion, FLUX.2 for photoreal + exact HEX brand colors — but reuse one prose-first master prompt across all three.
This is one lane of the full system. Get all ten — prompt skeletons, copy-paste templates, worked examples, and the 2026 tool picks — in The AI Creator's Playbook: get the complete 70-page playbook ▸