Here's a quick overview of what you'll get and the main steps. - Goal: keep one character looking the same across poses, styles and scenes. - Core tools: Midjourney in Discord, a strong reference image set, the --cref (character reference) method, fixed seeds and consistent model/version. - Time: plan 1–3 hours to build a base character and another hour to tune variations. - Legal/Affairs: paid Midjourney plans allow commercial use; if selling art in Canada and annual revenue exceeds $30,000 CAD register for GST/HST (see https://www.canada.ca/en/revenue-agency/services/tax/businesses/topics/gst-hst-businesses.html). - Quick links: Midjourney docs — https://www.midjourney.com/docs. Canadian copyright info — https://www.canada.ca/en/services/innovation/intellectual-property/copyright.html.
Prerequisites
Before starting, get these things in order.
- Midjourney account and Discord access. Midjourney runs in Discord — a free Discord account is enough to try things, but paid plans include higher limits and commercial rights. Check Midjourney's pricing and terms at https://www.midjourney.com.
- A set of 3–8 reference images of your character. Aim for different angles, lighting and outfits. Higher resolution is better — 1,000 px on the long edge is a sensible minimum.
- A consistent model/version choice. Midjourney updates models; keep the same model (for example v5 or v6) across your runs to avoid style drift.
- Basic prompt structure and a way to store prompts, seeds and image URLs (a spreadsheet or note app works fine).
- If you plan to sell work in Canada, know the GST/HST small supplier threshold: $30,000 CAD annual revenue. Register with the CRA if you cross that (https://www.canada.ca/en/revenue-agency/services/tax/businesses/topics/gst-hst-businesses.html).
Step-by-step: build a consistent character
Follow these steps in order. Number them, keep notes, and don’t change the model or seed unless you want variation.
-
Create a clean base image. Start by prompting Midjourney or uploading a hand-drawn / photographed reference. Keep the pose neutral and the face clear. Prompt example: “portrait of a young Canadian firefighter, neutral pose, high detail, cinematic lighting --ar 2:3 --v 5 --stylize 50”. Save the best result as your canonical base.
-
Upload reference images to Discord. Drag and drop your chosen base and angle images to a Midjourney bot channel. Right-click the uploaded image and choose “Copy image address” — you’ll use that URL in commands. Keep those URLs in a text file.
-
Use the --cref (character reference) command. Add the cref URL to your prompt so Midjourney knows which character to reuse. Syntax: add the image URL then append “--cref https://yourimage.jpg” at the end of the prompt. Example: “--cref https://cdn.discordapp.com/attachments/.../character01.jpg a full-body action shot of same character, running through a rainy city, cinematic lighting --ar 16:9 --seed 123456”.
-
Fix a seed and model version. Use --seed 12345 (choose any number) and include the model flag you used for the base (for example --v 5 or --v 6). The seed locks the image randomness and makes variations repeatable. Don’t change the model version between renders.
-
Make sure to keep the key prompt tokens consistent. Use the same key descriptors for face shape, hair, ethnicity, scars, clothing style and accessories — the AI treats these like attributes. For example: “short dark hair, olive skin, small nose, scar left eyebrow, leather jacket”. Put those phrases near the front of the prompt; they carry more weight.
-
Try using multiple reference angles. For better consistency upload 3–5 angles (front, 3/4, profile, back). Then include multiple --cref entries or a single composite reference URL that Midjourney can parse. This helps the system understand 3D structure and keeps features stable across poses.
-
Adjust the stylize and quality settings. Lower stylize values (for example --stylize 50 or --stylize 100) force the model to follow the prompt more literally. Use --quality 1 or --quality 2 for more detail. Higher stylize can change facial features, so keep it moderate for consistency.
-
Make a list of poses and scenes. Once you have the base, make a list of the specific scenes and angles you’ll generate — e.g., close-up angry, three-quarter smiling, full-body walking. Generate them one by one using the same cref, seed and model flags.
-
Create slight variations carefully. Use slight prompt changes for clothing, mood or lighting but keep the attribute block unchanged. If you want different outfits, add “outfit: red parka” rather than replacing descriptors about the face or hair.
-
Save your files with clear names and IDs. Save every result with a filename that includes model, seed, date and prompt snapshot — e.g., character01_v5_seed123456_2026-03-10.jpg. It’s faster to find the right image later and to re-run or upscale consistently.
-
Use upscalers or editing tools to clean up images. Use Midjourney’s U1–U4 upscalers or export the image to Photoshop/Procreate. Small edits — cleaning a mis-rendered eye or adjusting colour — keep the character consistent across final artwork.
-
Keep track of your prompts and settings. Keep a simple spreadsheet with the prompt, cref URL, seed, model, stylize and result link. This saves time if you need to reproduce or expand the set months later.
Tips
- Name features clearly, like “left-eyebrow-scar: yes,” so you remember. Short tags work well.
- Try to keep lighting consistent in your reference images because daylight and studio lighting can change how features look.
- If Midjourney adds unwanted details, use negative prompts (the --no parameter) — for example “--no glasses” or “--no text”.
- When you need the exact same outfit or item every time, include brand or material words — “worn brown leather jacket, brass zipper”. The AI responds to material cues.
- For animation or sequential frames, export character images and use them as reference frames for the next render — keep seed and cref identical.
Common mistakes to avoid
- Changing the model version mid-project. That’s the fastest way to break consistency.
- Using only one reference image. One angle invites hallucination when the character is posed differently.
- Relying on very high stylize values. They push the model to interpret rather than reproduce details — and faces shift.
- Not saving seeds and prompts. Without those you can’t reproduce earlier results.
- Ignoring copyright and commercial rules. Paid Midjourney subscriptions grant broader commercial use than free trials — check Midjourney’s terms and, in Canada, check Copyright Act guidance at https://www.canada.ca/en/services/innovation/intellectual-property/copyright.html. If selling art and revenue passes $30,000 CAD, register for GST/HST (https://www.canada.ca/en/revenue-agency/services/tax/businesses/topics/gst-hst-businesses.html).
Alternatives and quick comparison
Midjourney works well for stylized, painterly and cinematic characters. But there are other paths if you need different workflows:
- Stable Diffusion + DreamBooth / LoRA — great for tightly controlling a character through model fine-tuning on several images. Requires compute and some technical setup (local GPU or cloud). See Stability AI at https://stability.ai.
- ControlNet — useful when you want pose control (skeletons) while keeping a character’s visual identity.
- Custom commissioned art — still the best option if legal certainty and exact likeness are required.
Pick Midjourney when you want speed and strong visual style. Pick DreamBooth if you need pixel-for-pixel likeness and you can invest time and compute.
Related Articles
- The 4 Best Water Filter Pitchers (2026): PFAS, Microplastics
- Canada Remote Work Rights 2026 — Flexible Working Request Guide
Keeping a single character consistent in Midjourney comes down to planning — a solid base, multiple reference angles, fixed seeds and the same model. Use the --cref method, keep stylize low, and track every prompt and result. And if you plan to sell or publish work in Canada, check copyright rules and register for GST/HST once annual sales top $30,000 CAD. With a consistent setup you can produce a whole catalogue of poses, moods and scenes that read like the same person — whether it’s for comics, a game pitch, or cover art.
This article was created with AI assistance.
