Quick answer
Create a small set of “anchor” images (front, 3/4, close-up face, a neutral full-body) from a single base prompt, save and reuse those images as image-prompts, and keep a locked, explicit character-template prompt (traits, color palette, silhouette, lighting, camera, and style tokens). For ultimate repeatability, move to a fine-tuned model (DreamBooth/LoRA) when budget/quality needs justify it.

Why this works
Midjourney is easiest to steer with (A) strong, consistent textual descriptors and (B) an actual image reference you keep re-using. Seeds help a bit for composition/noise, but the reliable way to preserve the same face/body is to reference a canonical image or a fine-tuned model trained on that character.

Recommendation (one-line)
Use a hybrid workflow: produce and curate 4–6 anchor images in Midjourney, then reuse those anchor images plus a locked prompt template for every pose/outfit variation. Move to a DreamBooth/LoRA route only for large batches or commercial work that requires near-photoreal identity consistency.

Decision criteria (which path to choose)
- Use anchor-image + locked-prompt workflow if: you’re indie, need dozens (not thousands) of variations, and want fast iteration with low cost.
- Use fine-tuning (DreamBooth/LoRA) if: you need hundreds+ assets, pixel-perfect identity, or consistent character across multiple platforms (higher cost, requires compute/skill).
- Team/scale: small solo/team — anchor-image is simplest. Larger studio — fine-tune the model and integrate into a pipeline.

Practical checklist (step-by-step)
1) Create a base prompt template and lock it. Example template:
[ANCHOR_IMAGE_URL_1] [ANCHOR_IMAGE_URL_2] — "character name: tall young woman; narrow oval face; small upturned nose; pale freckled skin; ash-brown bob haircut with single blue streak; emerald eyes; slender athletic build; signature silver hoop earring on left; muted teal/patina color palette; soft cinematic rim lighting; 85mm portrait; painterly anime-realism style"
+ flags: --ar 2:3 --v 5 --q 2 --stylize 50 --seed 12345 --iw 1
2) Produce anchor images: run the base prompt and generate front, 3/4, close-up, and neutral full-body. Upscale and save the best results as your canonical references (download & host them where you can paste URLs into MJ).
3) Reuse anchor images for variations: upload the appropriate anchor images and add a short instruction for pose/outfit: e.g. "same character, wearing red leather coat and goggles, dynamic running pose, studio rim light" — keep the rest of the template identical.
4) Keep style and technical flags constant: same --v, --stylize, --ar, --q and same artist/style tokens. Small changes here cause big visual drift.
5) If results drift, iterate: generate a close match, save it, and use that new image as an anchor for tighter edits (the “best match as reference” trick).
6) Track metadata: keep a simple text file with the exact prompt, flags, anchor image filenames/URLs and seed used for each final.

Best-for / Avoid-if
- Best for: indie artists, concept shots, comics, game mockups where consistent look matters but perfect pixel-for-pixel identity isn’t critical.
- Avoid if: you need exact photographic likeness across hundreds of assets — then use DreamBooth/LoRA or a custom pipeline.

Extra tip
Use ChatGPT (or similar) to generate and normalize your locked prompt templates and variant lists (poses, outfits, props) so you don’t accidentally change wording between runs.

If you want, I can: (a) craft a reusable prompt template for your character from a short trait list, or (b) outline the DreamBooth/LoRA route and costs.

Compare Midjourney and Leonardo AI