How to get consistent characters in Midjourney workflow

Short answer: combine a high-quality reference sheet + a locked Midjourney model/flags + a reusable prompt template that always includes the same seed and image weights. Below is a step-by-step strategy, a ready-to-copy example prompt, decision criteria, and a practical checklist you can use every time you update a character.

Recommendation
Use one master reference image (or a set of 3: front, 3/4, profile) uploaded to Midjourney, lock the model version, and always regenerate variations from that reference using the same --seed, --ar, --stylize, and image-weight (--iw). Save the exact prompt, seed, and model in a small version-control document.

Step-by-step prompt & seed strategy
1) Create your reference sheet: one clean neutral-expression portrait at high resolution, plus hairstyles/costume close-ups. Keep background plain.
2) Upload the reference images to Midjourney and copy the image URLs (or use attachments in Discord). Use the best of these as your “character anchor.”
3) Use a prompt template that separates immutable traits (skin tone, eye shape, scar, hair color) from style/lighting. Example template (fill placeholders):
[refURL] [refURL2] "CharacterName: short unique tokens describing immutable features (e.g., scar on left eyebrow, deep-set green eyes, crescent mole under left lip)" cinematic portrait, 3/4 view, character design, realistic lighting, muted palette --ar 2:3 --v 5.2 --seed 12345 --stylize 50 --quality 2 --iw 1.0
4) Always use the same --seed number for new renders of this character. The seed fixes the noise pattern so features stay more consistent across renders.
5) Lock model/version (e.g., --v 5.2). Model updates change behavior; locking prevents drift.
6) Use image-weight (--iw) at 0.8–1.2 to prioritize the reference. Higher weight = closer to reference, lower = more creative variation.
7) For pose or outfit changes, keep the same reference + seed but add modifiers ("wearing leather jacket, holding map") and keep --no for unwanted changes (e.g., --no beard) to avoid drift.
8) When you need an exact face match for animation/sprites, produce several close-angle renders (same seed, different crops) and pick the best; optionally refine in an editor for pixel-perfect use.

Decision criteria
- Budget: Paid Midjourney subscription recommended (higher quality, seeds behave more predictably). If budget is tight, accept more manual cleanup.
- Skill level: Prompt templating is low-to-medium difficulty; use ChatGPT to generate templates if you want automation.
- Workflow stage: For early concept, loosen stylize/iw; for final assets, tighten them and do manual edits.
- Team size: Solo devs can centralize prompt/seed logs; teams need a shared prompt sheet.

Best-for / Avoid-if
- Best for: keeping facial features consistent across iterative updates and variations. Great for concept-to-sprite workflows.
- Avoid if: you need absolute photoreal identity across many poses (then consider 3D scans/rigging or manual artist retouches).

Practical checklist (use every session)
- [ ] Reference images uploaded and URLs saved
- [ ] Prompt template filled with exact trait tokens
- [ ] Model/version and seed recorded (--v, --seed)
- [ ] Flags set: --ar, --stylize, --quality, --iw
- [ ] Negative prompts (--no ...) tested if needed
- [ ] Outputs saved and indexed with seed + prompt

If you want, I can generate a tailored prompt template for your specific character traits. For scripting or automating prompts, use ChatGPT to create and vary templates while keeping the core tokens and seed locked.

Compare Midjourney and Leonardo AI