A real lookbook tells a story across 6-12 frames: the same model, evolving looks, a palette that hangs together. AI defaults will hand you twelve different models in twelve different lighting setups, which is useless as a brand asset. This tutorial locks the face, freezes the palette, and rotates only the outfit and pose so the set reads as one shoot instead of twelve random images. The workflow below is verified against the tools as they ship in June 2026.
TL;DR
- Generate one canonical model portrait first, then reuse it as a reference image for every outfit. The whole set rides on that one image being clean.
- For face lock, use Midjourney V7 Omni Reference (
--oref+--ow, default weight 100, push to ~300-400 to hold a face), or Nano Banana Pro (Gemini 3 Pro Image), which keeps up to 5 people consistent and blends up to 14 reference images. - Freeze the brand palette in HEX before generating, and vary only three things per outfit: garment, pose, one styling accent. Backdrop, lens, and light stay fixed.
- Review all six winners in a grid at thumbnail size. If a frame reads as a different shoot, re-generate it rather than retouching it.
What this covers
A face-locked lookbook workflow built on three primitives: a canonical model portrait, a frozen brand palette, and a rotating outfit/pose slot. The face-lock step has three viable engines as of June 2026:
- Midjourney V7 with Omni Reference (
--oref). V7 retired the old--crefcharacter-reference flag and replaced it with Omni Reference, which embeds any element (a person, an outfit, a prop) from a reference image and exposes a weight slider via--ow. - Nano Banana Pro, Google’s Gemini 3 Pro Image model. It edits an existing image while keeping the subject’s identity, supports 2K and 4K output, and renders text-on-garment legibly, which matters for branded pieces.
- Stable Diffusion (SDXL) with IP-Adapter FaceID Plus v2, for teams that need full local control and a repeatable seed.
Tool comparison (June 2026)
| Tool | Face-lock mechanism | Entry price | Notable for lookbooks |
|---|---|---|---|
| Midjourney V7 | Omni Reference --oref + --ow (1-1000, default 100) | $10/mo Basic (annual $8) | Strongest stylized-fashion aesthetic; fast iteration |
| Nano Banana Pro (Gemini 3 Pro Image) | Edit-from-reference, up to 5 people / 14 input images | In Gemini app on Google AI Pro $19.99/mo; also API | 4K output, accurate text on garments, localized edits |
| Stable Diffusion (SDXL) | IP-Adapter FaceID Plus v2, weight 0.6-0.8 + ControlNet | Free (self-hosted) / GPU cost | Full seed/pose control; near-100% lock with a trained LoRA |
Prices and version numbers above are current as of June 2026 and shift often; confirm on each vendor’s page before you commit a budget.
Who this is for
Indie fashion brands prototyping a season drop without a studio, e-commerce sellers staging six outfits for product pages, stylists pitching mood for client work, and creators building Instagram carousels that need to feel like one shoot.
When to reach for it
Seasonal lookbook concepts, Instagram carousel campaigns, landing-page hero rotations for a clothing line, pitch decks for stylists or brand-builders, and any case where you need 6-12 outfit variants on the same face in under a day.
Before you start
- Pick the model once. Stylized photoreal (a slight painterly hand) holds consistency far better than pure photoreal, so choose accordingly.
- Lock the brand palette in HEX before any generation: one neutral, one accent, one mood color. Every outfit must live inside it.
- Decide the shoot location and backdrop early: studio cyc, urban rooftop, gallery wall. Backdrop is the second-strongest signal of “one shoot”.
- Pick the deliverable aspect ratio up front (4:5 for IG feed, 3:4 for a lookbook PDF, 9:16 for Stories) and stay there. Midjourney uses
--ar 4:5; Nano Banana Pro takes the ratio in the prompt.
Step by step
- Generate the canonical model portrait. Front-facing, neutral expression, three-quarter body, even soft light. Spend 30-45 minutes here; the rest of the set rides on it. Save the final image URL (Midjourney) or file (Nano Banana / SD).
- Write the frozen model block. List 5-7 visible traits: hair color, length and texture; eye color; skin tone; body type; one signature mark. Paste this byte-identical into every outfit prompt. Rephrasing it is the number-one cause of drift.
- Lock the palette in words and HEX. Example:
warm sand, deep olive, ivory; HEX FBE7C6 / 545A2C / F4F2EA. Add this exact line to every prompt. - Vary only three things per outfit: garment description, pose, and one styling accent. Backdrop, lens, and light stay frozen across the set.
- Apply face lock for your chosen engine:
- Midjourney V7: drag the canonical into the Omni Reference bin (web) or append
--oref [canonical-url] --ow 350. Start around--ow 300-400for a held face; values above 400 get unpredictable unless you also raise--stylize. - Nano Banana Pro: upload the canonical, then instruct “keep this exact model and face, change the outfit to [X], keep the backdrop and lighting identical.” It edits rather than re-generating, which preserves identity well.
- Stable Diffusion (SDXL): IP-Adapter FaceID Plus v2 at weight 0.6-0.8, plus ControlNet OpenPose for the pose, then ADetailer to redraw the face cleanly.
- Midjourney V7: drag the canonical into the Omni Reference bin (web) or append
- Generate 4-6 takes per outfit. Reject anything where the face drifts, the palette breaks, or the backdrop shifts.
- Grid review. Open all six winners at thumbnail size. Anything that does not read as the same shoot gets re-generated, not edited.
Quality check
- Same face across all frames at thumbnail size? Run the squint test before the zoom test.
- Does the palette hold? Sample one pixel from each image; the dominant colors should map back to your three HEX codes.
- Is the backdrop consistent in lighting direction and color temperature? A mixed warm/cool set reads as two shoots.
- Garment details accurate? AI invents buttons and seams freely; check against your reference garment if you have one. Nano Banana Pro’s localized editing is the cleanest way to fix a single wrong detail without touching the face.
Reuse the workflow next season
- Save the model block, palette block, and backdrop block as a
lookbook-template.md. Next season, swap only the palette and outfits. - Build a library of outfit prompt blocks by category (outerwear, evening, athleisure) so you can compose a full season in about an hour.
- Keep the canonical portrait in a version-controlled
models/folder. When the model “ages” or pivots style, cut an explicit new version rather than drifting. - Re-check your face-lock weight every few weeks. Model updates can shift
--owbehavior, so a setting that held last month may need a nudge.
Common mistakes
- Rephrasing the model description per outfit. “Auburn hair” becomes “ginger” becomes “red”, and you have a different person by outfit four.
- Letting the backdrop drift. Six different backdrops look like six different shoots, even on the same face.
- Declaring the palette in vibe words only (“earthy”). Always include HEX or named pigments; vibe drifts, HEX does not.
- Chasing hyperreal photoreal. Small face shifts read as a different model; stylized photoreal is more forgiving.
- Mixing Midjourney faces with SD faces in one set. Pick one engine per lookbook and stay there.
- Skipping the grid review. Frame by frame, everything looks fine; tiled together is where breakage shows.
FAQ
- Does
--crefstill work in Midjourney?: No. Midjourney V7 retired--crefin favor of Omni Reference (--orefwith the--owweight). If you are following an older tutorial that uses--cref, switch to--orefand start the weight around 300-400 for faces. - Which tool keeps a face most consistent across outfits?: For pure ease, Nano Banana Pro edits from a reference and holds identity for up to 5 people. For aesthetic control, Midjourney V7 Omni Reference is strong. For maximum lock, a trained Stable Diffusion LoRA (15-30 images of the model) gets near-100% but takes setup time.
- Can I use a real model’s face?: Not without explicit written consent and a license. AI replication of identifiable real people is legally and ethically risky, and several platforms prohibit it.
- Why does the face still drift even at a high weight?: The canonical itself is probably inconsistent (multiple angles, varied lighting). Regenerate it to a cleaner front-on shot, then re-reference. A noisy reference image breaks face lock no matter how high you push the weight.
- Should the model change pose between outfits?: Yes, slightly. Identical poses read as paper-doll dress-up. Vary the pose within one vocabulary (standing, slight lean, hand on hip).
- How many outfits before drift compounds?: About 8-10 in one session is the safe ceiling. Beyond that, save the canonical and start a fresh session from the same reference.