3.5Character and series consistency
The essentials in 30 seconds
- The real pro challenge: one beautiful image is easy, ten consistent images are hard — each generation restarts from random noise, seed alone is not enough.
- Show, don't describe: --cref (character, + --cw) for the face, --sref for style — combinable for the same hero in the same world.
- Method + humility: one excellent master image, everything constant except the situation, documented recipe — and consistency very good but never absolute.
Running mini-case — Sonia creates a juice-brand mascot
Sonia, freelance graphic designer in Algiers, must illustrate each flavor of her client with the same little mascot — a stylized desert fox. Descriptive prompts alone give “cousins”: a different muzzle with each flavor. She changes method: she freezes ONE truly successful master image, reuses it with --cref for each new scene (changing only the situation), adds a common --sref for homogeneity, and documents version, parameters and recipe. Six flavors, one recognizable fox — with micro-retouches on two visuals. The client holds its reusable mascot.
Last episode of the Sonia thread — from light to style to the consistent series.
1. The real professional challenge: ten images, not one
Producing ONE beautiful image is easy; producing ten consistent images — same character, same world, same style — is the real professional challenge: illustration series, recurring character, storyboard, serialized campaign. The problem is structural: each generation restarts from random noise, so nothing guarantees continuity. Historically, people tinkered with seed (2.2) and very detailed prompts — a fragile method giving “cousins” rather than the same individual. Sonia lived it with her fox.
The lesson: textual description, however precise, does not freeze an identity. You must show the model who the character is, not just describe it.
2. Image references: --cref, --sref and their cousins
The modern answer: image references. In Midjourney, character reference (--cref, + weight --cw setting the resemblance strength) keeps the same face; style reference (--sref) keeps the visual signature (palette, texture, treatment). By combining them: same character, same world, different situations. On the Stable Diffusion side: IP-Adapter (face/style transfer) and LoRA (training a character/style — 4.3), even more robust.
Three-rule method. 1) First set an excellent reference image — everything depends on it, do not botch this step (several views: front, three-quarter). 2) Keep constant base prompt, model version and parameters; vary only the situation. 3) Accept imperfection: consistency is very good but not absolute — plan retouching (4.4) for demanding uses, and a simple graphic style (3.4) that replicates better than photorealism.
🔎 Concrete example — A reusable mascot, with method
Sonia serializes her fox across six flavors without letting it drift.
Validated master image → --cref for each new scene (only situation changed).
+ common --sref for graphic homogeneity.
→ simple style (3.4) + documentation (reference, version, parameters).Reading: one world, one identifiable character across six visuals.
Lesson: with no documented recipe, resuming the series is impossible.
Section vocabulary
- Series consistency
- Ability to produce several images sharing the same character, style or world.
Simple example: Sonia's six flavors with the same fox.
What it is not: six beautiful images with no link between them.
Reflex: target it from the brief, not afterwards.
- --cref (character reference)
- Reference image keeping the same face/character from one generation to the next.
Simple example: The fox master image reused for each flavor.
What it is not: textual description alone, giving “cousins”.
Reflex: show the character instead of describing it.
- --cw (character weight)
- Weight setting the resemblance strength to the reference character.
Simple example: The weight increased when the fox drifted too much.
What it is not: leaving the default weight when staging requires it.
Reflex: dose it: resemblance vs staging freedom.
- --sref (style reference)
- Reference image passing on the visual signature (palette, texture, treatment).
Simple example: The common --sref unifying the six visuals.
What it is not: changing style with each series image.
Reflex: combine it with --cref for character + world.
- IP-Adapter / LoRA
- Stable Diffusion-side equivalents: face/style transfer, or training a character or style (4.3).
Simple example: The trained LoRA if the fox becomes the brand's central asset.
What it is not: restarting each series from scratch with no capitalization.
Reflex: move to LoRA when the character becomes strategic.
In practice — Producing a consistent series
- Generate and select ONE truly successful character reference image (it conditions the whole series).
- Reuse it as character reference (--cref) with a suited weight, changing only the situation in the prompt.
- Add a common style reference (--sref) to homogenize palette and treatment across the series.
- Keep the model version, parameters and prompt formula constant; document everything to resume the series.
Ready-to-say script — handing the mascot recipe to the client (30 s)
“Here is your mascot and its recipe: master image, base prompt, model version, parameters and references. Each new flavor will restart from this base — same character, same style. Keep this folder: without it, resuming the series in six months is impossible.”
To say on delivery; hand over the recipe folder with the files.
Copy-paste template — consistent-series prompts
[master-image-url] little fennec fox mascot, flat vector, orange and teal palette --cref [url-ref] --cw 80 --sref [url-style] --ar 4:5
same base, the mascot holding an orange juice bottle, market stall background --cref [url-ref] --cw 80 --sref [url-style] --ar 4:5
same base, the mascot running with a mint leaf, minimal background --cref [url-ref] --cw 80 --sref [url-style] --ar 1:1Freelancer
One master image + --cref + noted recipe: the method retaining a series client.
SME in Algiers
Sonia serializes the fox across six flavors, retouches two visuals, documents everything.
Project team
Archive master image, version and parameters in a shared folder before producing.
Beginner
This week: one excellent reference, three --cref variants, one recipe sheet.
✓ Best practices
- Craft the reference image: the rest of the series depends on it.
- Vary only the situation; keep version, parameters and base prompt constant.
- Document reference, version and parameters to resume the series later.
✗ Common pitfalls
- Botching the reference image, then wondering why the whole series drifts.
- Varying everything at once (angle, light, style): drift explodes.
- Expecting absolute consistency: plan retouching for demanding uses.
Key takeaways
- Series consistency is the real professional challenge: each generation restarts from randomness, nothing is spontaneously consistent.
- Modern solution: image references — --cref (character, with --cw) and --sref (style), combinable.
- Method: one excellent reference image, then keep EVERYTHING constant except the situation; document the series.
- Consistency is very good but never absolute: plan retouching for demanding uses.
48h challenge — apply tomorrow
Freeze 1 master image, serialize 3 --cref + --sref situations and write the recipe sheet (version, parameters, references).
- My character starts from a master image, not a description.
- I combine --cref (+ --cw) and --sref, everything else constant.
- My recipe is documented and my series retouched if needed.
Frequently asked questions
Can I create a durably reusable brand character (mascot)?
Yes, with method, in four steps. (1) CREATE and freeze the reference: explore until a fully convincing image — your master image, ideally in several views. (2) DOCUMENT the recipe: prompt, version, parameters, seed, reference file — without it, resuming the series in six months is impossible. (3) SERIALIZE with --cref (only situation changed) + common --sref, adjusting the weight. (4) Build a LIBRARY of validated visuals. Reality check: very good but imperfect consistency (retouching to plan, 4.4); if the character becomes central, consider a LoRA (4.3). Tip: favor a SIMPLE graphic style (3.4).
Why does my character slightly change with each image despite the reference?
Expected behavior: the reference GUIDES toward a resembling face, but the image is regenerated from new noise — the model does not know your character as a fixed identity. To REDUCE drift: raise the WEIGHT (--cw), LIMIT variations (one thing at a time), REINFORCE the distinctive-trait description, keep EVERYTHING constant, PICK a good reference (sharp, lit, front/three-quarter). Otherwise: Photoshop retouching (4.4) or trained LoRA (4.3).
How do you keep the same character from one image to the next in Midjourney?