When you need a series of images — for an article, a product line, a personal brand, or a course — the worst part of AI generation is that 8 images look like 8 different artists made them. Different color, different brushwork, faces that don’t match.
Fastest fix (solves ~70% of cases): reuse one fixed “style header” prompt on every image and lock the seed. Everything below is the full toolkit, in ROI order, with the current Midjourney V8.1 and gpt-image-2 syntax.
Which bucket are you in?
| Symptom | Most likely cause | Go to |
|---|---|---|
| Every image a different color/mood | No fixed style block; color drift | Method 1 + 4 |
| Same character looks like siblings, not one person | No identity lock | Method 3 (--oref) |
| Overall look drifts shot to shot | Random seed + loose style | Method 1 + 2 |
| Need pixel-stable character for a long project | Prompting alone won’t hold | Method 5 (LoRA / Flux Kontext) |
| Two images use two different models/versions | Apples to oranges | See “When it isn’t your prompt’s fault” |
Why series consistency is hard
Each generation is a random sample from a distribution. Even with an identical prompt, a different seed — or a different model version — gives a different result. To make a series feel like one artist made it, you have to manually lock down a few variables. You can’t wish it consistent; you pin it.
6 methods to make a series consistent
In ROI order (highest to lowest impact per minute spent).
1. Reuse a fixed “style header” prompt
Split every prompt into two pieces.
Style header (constant):
clean cel-shading anime, soft pastel palette, KyoAni style, gentle warm light, 50mm portrait, painterly background --ar 4:5Subject (variable):
A girl reading a book by the windowCopy the style header verbatim into every prompt; change only the subject. Results immediately look like one artist’s work. This is the cheapest fix with the biggest impact, so do it first even if you also do everything else.
2. Lock the seed
- Midjourney: add
--seed 12345to your prompt. To recover the seed of an image you already like, react to it with the envelope emoji in Discord (or read it from the job page on the web app); the bot DMs you the seed. - Stable Diffusion / Flux (Automatic1111, ComfyUI, Forge): set the seed field instead of leaving it at
-1(random). - gpt-image-2 / ChatGPT: no public seed control — use Method 6 (reference chaining) instead.
Same seed + similar prompt = high resemblance. Keep the seed fixed and vary only the subject portion of the prompt.
Note: a locked seed nudges results closer together but does not freeze them. In Midjourney V8.1 the seed is a starting point, not a guarantee, so pair it with a style header or reference.
3. Use a reference image (style and/or character)
Midjourney updated its reference system across V7/V8.1, and the old --cref flag is now V6-only. Use the right tool for what you’re locking:
- Lock the look (color, medium, lighting):
--sref <image-url-or-code> --sw 100.--sw(style weight) ranges 0–1000, default 100; raise it to push the reference style harder, lower it to let your text lead. You can pass a numeric sref code instead of an image to reuse the exact same style across sessions. - Lock the character (same person across scenes):
--oref <image-url> --ow 100(Omni Reference, V7/V8.1).--ow(omni weight) ranges 1–1000, default 100. As of June 2026, keep--owbelow ~400 for predictable results; use 0–50 when you want the character’s essence but a big style change, and 400+ to tightly replicate facial features.--crefstill works only on V6/Niji V6. - Both at once: pass the same image as
--oref(content) and--sref(look) to lock identity and style together. - DALL-E / gpt-image-2: no
--sref; upload your first image as a reference and ask for “the same character/style as the attached image.” With Thinking mode on (Plus/Pro/Business), gpt-image-2 can return up to 8 coherent images in one prompt. - Stable Diffusion / Flux: IP-Adapter or ControlNet (reference/style) from an uploaded reference.
A reference strength around 30–60% of the maximum usually balances fidelity against variety.
4. Constrain the color palette
About 90% of perceived “style drift” is color drift. Spell out the palette in the prompt:
palette: warm cream, dusty pink, faded teal, soft brownOr pin exact hexes:
color palette limited to #f4e3c1 #d29b7c #8aa1a4 #4a3a30The model biases toward those colors and the whole series feels unified. Put the palette near the end of the prompt and avoid conflicting words like vibrant or saturated, which override it.
5. Train or use a LoRA (or Flux Kontext)
For long-term projects (one recurring character or a signature style), prompting alone won’t hold. Two durable options:
- LoRA (Stable Diffusion / Flux): train your own, or grab a ready-made one from Civitai or Hugging Face.
- Character LoRA: 15–25 images of one person from multiple angles.
- Style LoRA: 20–50 images of one consistent style (Flux trains decent style LoRAs on 20–30 images).
- Once trained, every generation is consistent by default.
- Flux.1 Kontext: an edit-and-generate model built for consistency — it preserves a reference character or object across new scenes and survives multiple successive edits with minimal drift. Good when you want consistency without training a LoRA.
6. Image-to-image chaining
Generate image 1 freely. For every later image, run image-to-image with the previous output as the input, plus a new subject prompt. Each image inherits the brushwork and color of the last.
- Image 1: free generation.
- Image 2 onward: previous image as input, strength 60–70%, new subject.
- The whole series stays close to the original look.
This is also the main consistency lever in tools without seed control (DALL-E / gpt-image-2): feed back the previous good image every time.
Shortest path
In effort order:
- Extract a style header prompt — 30 seconds
- Lock the seed — 1 minute
- Add palette constraints — 1 minute
- Add a reference image (
--sref/--oref) — 5 minutes - Train a LoRA — 1–2 hours, but solves it permanently
If you only do one thing: style header + seed lock. Together they handle about 70% of consistency problems.
How to confirm it’s fixed
- Lay 4–6 of the new images side by side at thumbnail size. If they read as one set at a glance, you’re done.
- Eyedrop the dominant colors of two images. If the palettes match within a few shades, color drift is fixed.
- For characters, line up the faces: same hair color, eye color, and face shape across shots means your
--oref/ LoRA is holding. - If one image still stands out, regenerate just that one with the same seed/reference before changing the whole pipeline.
When it isn’t your prompt’s fault
- Model version changed — e.g. Midjourney V8.1 became the default on June 10, 2026 (replacing V7), and V7 → V8.1 shifts the look. Pin the version explicitly with
--v 8.1(or whichever you started on) so a silent default change doesn’t re-style your series. - You’re actually using different models — one Midjourney image plus one Stable Diffusion image will never match. Pick one engine per series.
- Your reference image is itself stylistically messy — garbage in, garbage out.
- Same prompt, different account/hardware — minor model updates cause small variations; lock the version flag.
Easy misjudgments
- “My prompt isn’t specific enough.” Sometimes very detailed prompts still drift; the missing piece is usually the seed or the style header, not more adjectives.
- “The model can’t do it.” If a colleague gets a consistent series on the same model, it’s method, not model.
- “Only the faces are inconsistent.” Usually it’s color and brushwork drift; faces just make it obvious. Fix the palette and style header first.
- “LoRA is too complex.” A character LoRA is surprisingly approachable — roughly 30 minutes to train on a service like Civitai.
Prevention
- For any series project, lock the style header template on day one.
- Keep your seed / sref / oref / palette in a single project doc.
- Save the original prompt for every image (Midjourney stores it on the job page; local tools can read it from PNG metadata / EXIF).
- Pin the model version flag (
--v ...) so a default upgrade doesn’t restyle mid-project. - Long-term projects: just train a LoRA up front.
FAQ
Q: How do --sref and --sw work in Midjourney?
A: --sref <image-url-or-code> sets the style reference; --sw 100 sets style weight (default 100, range 0–1000). Higher pushes the reference style harder; lower lets your text lead. Multiple references go after one --sref, separated by spaces.
Q: I locked the seed but faces still vary. What now?
A: Seed alone won’t hold a face. On Midjourney V7/V8.1, add Omni Reference: --oref <image-url> --ow 100 (raise --ow toward 300–400 for tighter facial replication). --cref only works on V6.
Q: How many images do I need to train a LoRA? A: Character LoRA: 15–25 multi-angle shots. Style LoRA: 20–50 same-style images (Flux does well on 20–30). Quality beats quantity — clean, varied, on-style images matter more than count.
Q: Can I get a consistent series with ChatGPT / DALL-E? A: Yes, and it’s stronger than before. As of June 2026, gpt-image-2 (ChatGPT Images 2.0) in Thinking mode can return up to 8 coherent images with a consistent character and style from one prompt (Plus/Pro/Business). There’s still no public seed control, so for follow-up images, attach the previous good image as a reference.
Q: My palette keeps getting overridden by other prompt words. How do I lock it?
A: Put the palette: block at the end of the prompt and remove conflicting words (vibrant, saturated, colorful). On Midjourney you can also lower the stylize value (--s) so the model imposes less of its own color.
Q: Which Midjourney version should I use for a consistent series?
A: As of June 2026 the default is V8.1 (HD 2K, ~4–5x faster, --raw to strip default styling). Whichever version you start a series on, pin it with the version flag so a default change doesn’t re-style later shots.
Related articles
- How to Fix Distorted Faces in AI Images
- How to Improve Motion Consistency in AI Videos
- Best Anime Character Image Prompts
- Realistic Beautiful Woman Portrait Prompts
- ChatGPT prompt improvement
- Claude prompt best practices
- Refactor prompts
Tags: #Image generation #Consistency #Midjourney #Prompt #Debug