TL;DR
To make 6 AI images read as one series, you need three things, not one prompt: a single locked reference image, a master prompt template with a [SUBJECT] placeholder, and 3 explicitly quantified levers (lighting, palette, composition). Feed all three to a model that supports image-based style locking — Midjourney V7 (--sref for style, --oref for objects/characters), Nano Banana Pro, or FLUX.1 Kontext. Render at least 3 variations per slot and curate. A style anchor cuts variance by roughly 70-80%, not 100%; the rest is closed by curation.
The task
You’re shipping a 6-part blog series next month and need a cover image for each post. You sit down with Midjourney, write a careful prompt for the first one, and get a great editorial illustration. You write a similar prompt for the second — different subject, same style language — and back comes a 3D render in a completely different palette. By image 4 you have six perfectly fine images that look like they came from six different artists. A blog series needs visual consistency or readers will not register it as a series. You want a master prompt template plus one reference image that locks lighting, palette, and composition across all 6 outputs.
Where AI helps — and where it does not
AI image models carry inherent variance even with identical inputs: same prompt, same seed, two runs, slightly different output. A style anchor (one reference image + a master prompt template + 3 locked levers) reduces that variance by an estimated 70-80%, not 100%. That is the realistic ceiling. The remaining 20-30% of drift gets handled by rendering 3 variations per slot and curating to the best one.
What AI cannot do: tell you when a subject genuinely needs a style break. If five of your six subjects are abstract concepts and one is a specific real product, the product may need different composition rules. The model will not volunteer that — flag it yourself. AI also cannot hold consistency across model versions; if you switch from Midjourney V6 to V7 mid-series, or swap a model entirely, the style drifts regardless of prompts. As of June 2026, even Midjourney’s own --sref style codes shifted between versions — old codes need --sv 4 to behave like before, so re-pin your reference when you upgrade.
A specific failure mode: the model defaults to describing a style broadly (“editorial illustration, pastel palette”) without pinning the levers tightly enough to reproduce. Force it to name lighting direction, palette in hex or named colors, and a composition pattern explicitly — not just adjectives.
What to feed the AI
- A single locked style reference image (the series anchor) — your best version from a prior asset, or one you generate first and freeze
- The list of 6 subjects you need rendered, each in a short phrase
- Forbidden elements that must not appear in any image (text overlay, faces, specific brand objects, certain colors)
- The model you will use (Midjourney V7, Nano Banana Pro, FLUX.1 Kontext, Seedream — reference-image syntax differs by model)
- Aspect ratio and platform constraints (1200x630 OG card, 1080x1080 IG, 9:16 Reels)
- The platform the images appear on (blog, App Store, social, ad)
- Your seed strategy — locked seed across the series, or unique per subject (Midjourney supports both)
- One subject where the style applies to an unusual case — these stress-test the master prompt
Which model locks style best (June 2026)
| Model | How you pin style | Strength | Watch-out |
|---|---|---|---|
| Midjourney V7 | --sref [code/url] for style, --sw 0-1000 for strength; --oref [url] + --ow 1-1000 (default 100) for objects/characters | Editorial and illustrative consistency | --oref costs ~2x GPU time and is not compatible with Draft/Fast mode; style codes changed across versions |
| Nano Banana Pro (Gemini 3 Pro Image) | Upload reference image(s); describe style in natural language | Photoreal, legible in-image text, up to 4K | Character consistency is strong but not guaranteed every render |
| Nano Banana 2 (Gemini 3.1 Flash Image) | Same as Pro, faster | Holds up to 5 characters / 14 objects across a workflow | Flash speed trades some fidelity vs. Pro |
| FLUX.1 Kontext | Image + text input; preserves style across edits | Multi-turn editing without losing identity | Style transfer can need tuning; character lock is its strongest mode |
When the look is illustrative, Midjourney --sref is still the most reliable lock. When you need real product photos or readable text baked into the image, Nano Banana Pro pulls ahead. Test 1-2 references in two models before committing the whole series.
Copy-ready prompt
Generate a master prompt template + 6 final prompts for a series of images that share visual style.
Style reference image (link or detailed description): [paste]
Model I'll use: [Midjourney V7 / Nano Banana Pro / FLUX.1 Kontext / Seedream]
6 subjects to render: [paste]
Forbidden elements: [paste — e.g., "no text overlay, no human faces, no brand logos"]
Aspect ratio: [paste — e.g., 16:9, 1:1]
Seed strategy: [locked seed / unique per subject]
Return:
1) Master prompt template — one block with a [SUBJECT] placeholder, plus 3 explicit lever rules:
- Lighting (direction, hardness, source — not just "dramatic")
- Palette (3-5 specific colors with hex codes or precise named colors — not just "pastel")
- Composition (subject placement in frame, negative space %, foreground/background layering)
2) 6 final prompts — master prompt with each subject substituted into [SUBJECT].
3) Per-subject tweaks — for any subject that genuinely needs adjustment beyond the [SUBJECT] swap (e.g., "subject 4 is a product shot — adjust composition to 60/40 rather than centered"), flag it explicitly with the tweak.
4) Curate-and-render protocol — render 3 variations per subject (at least), and a checklist of what to look for to pick the best (palette match, composition adherence, prohibited elements absent).
Rules:
- Lever rules must be specific enough that someone else could reproduce the style. "Soft lighting" fails; "soft diffused side-lighting from camera left, 45° elevation, no hard shadows" passes.
- Palette must list specific colors. "Pastel" fails; "sage #B3C7B0, cream #F5EFE2, dusty pink #D4A5A0" passes.
- The master prompt should fit in under 60 words. Beyond that, model attention degrades.
- For any subject the master prompt cannot handle without modification, name the modification — don't pretend a one-size-fits-all prompt works across 6 different subjects.Shorter variant — refresh existing series
I have 6 images that mostly match but drift on palette. Below is the prompt I used and the reference image description. Tighten the palette spec to specific hex codes and rewrite the prompt to lock palette without changing lighting or composition.
Current prompt: [paste]. Reference image description: [paste].Sample output
A useful master prompt: “Editorial illustration of [SUBJECT], soft diffused side-lighting from camera left at 45° elevation with no hard shadows, palette of sage #B3C7B0 / cream #F5EFE2 / dusty pink #D4A5A0 / muted slate #6B7B8C, centered composition with subject occupying middle 60% and 20% negative space at top, foreground/background layered with subtle vertical separation, no text overlay, no human faces, no brand logos. Editorial publication style, 16:9 aspect ratio. --v 7 --style raw --sref [code] --seed 42”
A useful per-subject tweak: “Subject 4 is ‘AI cost dashboard’ — keep all lever rules but shift composition from centered to 60/40 left-weighted to fit the dashboard rectangle. Subject 6 is ‘meeting room scene’ — add ‘two figures, backs to camera, mid-frame, no facial detail visible’ to satisfy the forbidden-faces rule without removing the scene’s narrative.”
A useful curate-and-render note: “Render 3 variations per subject (Midjourney with --sref [code] --seed 42 keeps composition while giving you variation; for a unique-seed strategy, render 5 per subject). Curate criteria: (1) palette matches all 4 hex anchors within visible tolerance, (2) lighting direction is left-side, (3) no forbidden element appears, (4) subject is recognizable. Reject any image that fails any criterion — don’t ship a near-miss because the rest of the series is already done.”
How to refine
- Tighten palette and lighting before composition: “If the series is still drifting visually, rewrite the palette spec first (specific hex codes or named-color codes) and the lighting spec second (direction, hardness, source). Composition drift is the least visible to readers; palette drift is what they notice first.”
- Force lever specificity: “Re-read the master prompt. Any phrase like ‘soft pastel,’ ‘dramatic lighting,’ or ‘minimalist composition’ must be replaced with quantified specifics: a hex code, a degree of elevation, a percentage of negative space. Adjectives let the model improvise; specifics constrain it.”
- Add the right reference parameter: “On Midjourney V7, add
--sref [code]to lock the look and--swto dial its strength; add--oref [url] --ow 100only when a specific object or character must recur. On Nano Banana Pro or FLUX.1 Kontext, attach the reference image directly instead of a code.” - Vary composition per group, not per subject: “If all 6 subjects use the exact same composition the series looks rigid. Group subjects into 2 sets of 3, vary composition between sets (one centered, one 60/40), but lock palette and lighting across all 6. Variation within a system reads as ‘series’; variation across rules reads as ‘unrelated.’”
- Render 3 per slot minimum: “Add to the protocol: render at least 3 variations per subject and curate to the one closest to the reference. Single-render workflows look identical in theory but drift 20-30% in practice; curation closes the gap.”
Common mistakes
- Trying to lock style with words only — pair the prompt with a reference image where the tool supports it (Midjourney
--sref/--oref, Nano Banana Pro reference upload, FLUX.1 Kontext image input, Seedream style transfer) - Asking for “consistent style” without naming the 3 levers — lighting, palette, composition are the levers; without specifying them you are hoping the model picks the same instinct every time, and it will not
- Generating one image per subject — you need 3 minimum and a human curate step; single-render workflows produce drift you only see when all 6 sit side by side
- Switching models mid-series — Midjourney V6 to V7, or Midjourney to FLUX, produces visibly different styles regardless of prompts; lock the model at the start
- Using adjectives instead of quantified specifics — “soft pastel” is interpretive; “sage #B3C7B0 + cream #F5EFE2” is enforceable
- Forgetting the forbidden-elements list — every series has 1-2 things that must never appear (text, faces, competitor product cameos); if you don’t list them, they appear
- Letting the master prompt grow past 60 words — model attention degrades; tighten or split into a base prompt plus per-subject additions
- Not rendering a stress-test subject early — if subject 1 is “tree” and subject 6 is “spreadsheet,” render subject 6 second to see whether the style breaks on an unusual subject, then adjust before you have burned 4 generations
FAQ
- Which models support reference-image style locking best in 2026?: For editorial and illustrative styles, Midjourney V7
--srefis the most reliable lock, with--oreffor recurring objects or characters. For photoreal work and legible in-image text, Nano Banana Pro (Gemini 3 Pro Image) leads, and Nano Banana 2 adds Flash speed plus consistency across up to 5 characters / 14 objects. FLUX.1 Kontext is strongest for multi-turn editing where identity must survive several edits. Test 1-2 references in two models before committing. - What is the difference between
--srefand--orefon Midjourney V7?:--srefpins the overall style (lighting feel, palette, rendering);--sw 0-1000sets how hard it applies.--oref(Omni Reference) pins a specific object, character, or logo so it recurs, with--ow 1-1000(default 100) for strength. Note--orefcosts roughly 2x GPU time and is not compatible with Draft or Fast mode. For a style series,--srefis the primary tool; reach for--orefonly when an exact element must repeat. - Should I add a seed number?: On Midjourney, yes — when the rest of the prompt is locked. Seeds let you tweak word-level prompts without losing composition. A locked seed across the series gives compositional consistency; a unique seed per subject gives variation. Choose based on whether you want the series to feel “uniform” (locked) or “related but varied” (unique).
- What if one subject just doesn’t work with the style?: Flag it as a per-subject tweak. If the master prompt cannot handle a specific subject (say, a product shot mixed with editorial illustrations), the honest move is a composition adjustment for that subject — not abandoning the style. Document the tweak so the team knows it was deliberate.
- How do I keep style across a model-version upgrade mid-project?: Don’t. Lock the model and version at the start of the series and finish before upgrading. If you must upgrade mid-series, plan to regenerate all prior images on the new version, not just the new ones — Midjourney’s style codes alone shifted enough between versions to break a series.
- The model keeps producing different palettes — what changes?: Add: “Palette is the highest-priority lever. Use the exact hex codes I specified. If the rendered image doesn’t visibly match all 4 hex anchors, reject and re-render. Treat palette as a constraint, not a suggestion.”
Related
- AI moodboard prompt
- AI brand style prompt
- Prompt Template for Consistent Image Style Across an Entire Series
- AI Sets Product Visual Direction
- AI Product Advertising Image Prompts
For the current parameter list, see Midjourney’s official Omni Reference docs.
Tags: #AI writing #Marketing #Workflow #Consistency #Image prompt