The hardest part of AI image generation isn’t getting one good image — it’s getting ten that feel like they belong together: same lighting, same palette, same mood, same camera language. A reusable prompt block does most of the work, and the newest model features (Midjourney’s style codes, FLUX.2’s reference stack, GPT Image 2’s multi-panel mode) do the rest. Here’s the template plus the exact settings that hold it together as of June 2026.
TL;DR
- Split every prompt into four layers. Lock the style/color/lighting layers word-for-word; vary only subject and composition.
- On Midjourney V7, add a frozen
--srefstyle code (and--oref+--owfor a recurring character).--crefis gone in V7. - On FLUX.2, attach the same reference image (up to 10) to every generation and pass hex color codes directly.
- On GPT Image 2, use one prompt that asks for a multi-panel sheet so the model holds style across panels in a single pass.
The 4-layer prompt structure
[Subject] // what's in the frame
+ [Style] // the visual DNA — locked across the whole series
+ [Lighting & color] // also locked across the series
+ [Camera / composition] // varies per image to add varietyThe trick: lock the middle two layers across every prompt. That’s what produces visual consistency.
Master style block (paste this into every prompt)
Style: editorial photography, slightly desaturated film look, soft analog grain, Kodak Portra 400 emulation
Color: warm muted palette, dusty rose + sage green + cream, low contrast
Lighting: soft window light from camera-left, gentle falloff, no hard shadowsThis is your “brand sheet.” Reuse it word-for-word across every image in the series.
Then vary the subject + composition
For image 1:
A young woman writing in a notebook on a wooden desk, half profile, --ar 4:5
[paste master style block here]
Composition: medium shot, shallow depth of field, subject slightly off-centerFor image 2:
The same desk, now with a cup of coffee and an open laptop, no person, --ar 16:9
[paste master style block here]
Composition: top-down flat lay, balanced negative spaceNotice how the style/lighting/color block stays identical. That’s what carries consistency.
Lock style with model-native features (June 2026)
The text block sets intent; the model’s reference features enforce it. Per-tool settings that matter right now:
| Tool | Lock style with | Lock a character with | Reference limit / note |
|---|---|---|---|
| Midjourney V7 | --sref <code> (style code), --sv 6 default | --oref <url> --ow 100 (Omni Reference) | --cref no longer works in V7; Omni Reference costs 2x GPU time |
| FLUX.2 | Same reference image on every gen + hex color params | Up to 10 reference images per request | Released Nov 2025; native 4K, accepts hex codes like #C9A directly |
| GPT Image 2 | One prompt asking for an N-panel sheet | ”same character” in the same prompt | Released Apr 21, 2026; “thinking mode” holds 8 panels in one pass |
Source: Midjourney Omni Reference docs and Black Forest Labs FLUX.2.
Midjourney V7 specifics
--sref copies overall style — color, medium, texture, lighting — and is compatible with V6 and V7. Generate one image you like, grab its style code, then append the same --sref code to every prompt in the series. There are six style-reference variants in V7; pick one with --sv and keep it fixed (--sv 6 is the default).
For a recurring person, --cref is deprecated and silently ignored in V7 — use Omni Reference instead: --oref <image-url> --ow 100. The --ow weight runs 1–1000 (default 100). Keep it below 400 unless you’re using a very high --stylize, or results get unpredictable; ~400 is good for preserving a face and clothing details, ~25 for loose style transfer. Omni Reference costs 2x the GPU time of a normal V7 image, so budget Fast hours accordingly.
FLUX.2 specifics
FLUX.2 (Black Forest Labs, released November 2025) accepts up to 10 reference images in a single request and keeps identity consistent across outputs — useful when you need the same face across many backgrounds. It also takes hex color codes as direct parameters (e.g. #D9A5A0), so put your palette hex values in the prompt instead of trusting the model to interpret “dusty rose.”
GPT Image 2 specifics
GPT Image 2 (gpt-image-2, released April 21, 2026, built on the GPT-5.4 backbone) can render a multi-panel sheet — up to 8 panels — from a single prompt with consistent characters, object placement, and palette. For a small series, ask for the whole sheet at once (“an 8-panel sheet, same character and same color grade in every panel, varying only the pose”); the model holds consistency better in one pass than across separate generations.
Character consistency
For series featuring the same person, pin a character block in the text and use the model feature above:
Character: a 30-year-old Asian woman, shoulder-length dark hair, round wire-frame glasses, oversized cream sweater, soft natural makeup
[paste master style block here]
Scene: [varies per image]Pin the character block. Vary only the scene. On Midjourney V7, also feed a reference portrait via --oref so the face survives pose and lighting changes; on FLUX.2, attach the same reference image to every generation.
Industry-specific style blocks
Product photography:
Style: clean studio product photography, seamless paper background, minimal
Color: neutral grays, single accent color matching the product
Lighting: large softbox top-left, fill card right, subtle reflection underneathEditorial illustration:
Style: contemporary editorial illustration, flat shapes with subtle gradients, hand-drawn texture overlay
Color: limited palette of 4 colors max, off-white background
Lighting: implied direction, soft shadows onlyCinematic photography:
Style: anamorphic cinematic look, slight lens flare, 35mm film aesthetic
Color: teal and orange grade, deep shadows, milky highlights
Lighting: practical sources only (window, lamp), motivated directionPro tips that actually move the needle
- Reuse exact words, not synonyms. “Cinematic” ≠ “filmic” to the model.
- Save the master block and your
--srefcode in a sticky note. Don’t rewrite either from memory. - Test 3 images first. Then lock the block (and the style code). Then generate the rest.
- Pin a seed. Midjourney
--seed, or a fixed seed on FLUX, removes one more variable on top of everything above.
Cost note
This is a Fast-hours game on Midjourney, since you’ll iterate. As of June 2026, Midjourney is Basic $10, Standard $30, Pro $60, Mega $120 per month (20% off on annual; no free tier), and Omni Reference burns GPU time at 2x — Standard’s 15 Fast hours go quickly on a character series. See the official Midjourney plan comparison.
FAQ
Why does my series still drift even with the same style words?
The text block is intent, not a lock. Add a model-native anchor: a frozen --sref code on Midjourney V7, the same reference image on FLUX.2, or a single multi-panel prompt on GPT Image 2. Synonym swaps (“filmic” vs “cinematic”) and a missing seed are the two most common drift causes.
Does --cref still work in Midjourney?
Not in V7. --cref is incompatible and is ignored or errors. Use Omni Reference (--oref <url> --ow 100) instead; --ow ranges 1–1000 and should stay under 400 for predictable results.
Which tool is best for a 10-image series with one recurring character?
FLUX.2 if you have a clean reference portrait — it takes up to 10 reference images and holds identity well across backgrounds. GPT Image 2 if you want all panels in one pass (up to 8). Midjourney V7 if you care most about aesthetic polish and are happy to feed an --oref portrait.
Can I use hex colors instead of describing the palette?
On FLUX.2, yes — it accepts hex codes (e.g. #D9A5A0) as direct parameters, which is more reliable than “dusty rose.” Midjourney and GPT Image 2 still parse natural-language color, so keep your palette words identical across prompts.
How much will a full series cost on Midjourney? Plan by Fast GPU hours. Standard ($30/mo) includes 15 Fast hours; Omni Reference doubles GPU cost per image, so a 30-image character series with iterations can eat a meaningful chunk of that. Relax Mode (Standard and up) is unlimited but slower.
Related
- Style consistency across images — production tutorial for the same workflow
- AI image style drift fix — recover when the series drifts mid-batch
- AI consistent character images — character-side consistency
- Game character portrait sheets — applied example of the same approach
- AI Locks Visual Style Across a Series of Images