Midjourney reference images: build a reusable shot list before you prompt

Reference images are more useful when each one has a job. Do not throw ten images at the model and hope it understands your taste.

The fastest way to waste image generations is to start with a vague mood board. You add a portrait, a color palette, a product photo, a cinematic still, and an interior design image. The model reads them all as visual pressure, not as a structured brief. The output becomes an average of everything and a match for nothing.

This workflow turns reference images into a small shot list. It works for Midjourney and transfers to other image tools that support image references.

What you need before prompting

Prepare three reference types:

• Composition reference: where objects sit in the frame
• Mood reference: lighting, color, atmosphere, and contrast
• Material reference: texture, surface, clothing, product detail, or environment

The point is not to match any one image exactly. The point is to separate visual decisions so the prompt can be clear.

Step 1 - name the output format

Start with the thing you are making, not the tool command. A good shot list begins like this:

Output: square editorial hero image
Use case: website article card
Subject: person at a desk reviewing AI-generated image variations
Style: warm editorial photography, paper texture, not glossy SaaS

This gives you a standard to judge the result. Without it, every image that looks "cool" can distract you.

Step 2 - assign one job to each reference

Create a small table:

The "what not to copy" column matters. Image models are good at imitation and bad at knowing which part of a reference you meant. Spell it out.

Step 3 - write the base prompt

The base prompt should be shorter than you think:

Editorial photograph for a practical AI tutorial website. A focused solo creator sits at a wooden desk comparing AI image variations on a laptop. Warm side light, paper notebooks, realistic workspace, muted rust and ink palette, natural texture, calm magazine feel. Avoid glossy SaaS look, neon lighting, fake UI, unreadable text.

Notice what is missing: no long list of artists, no fake camera model unless it matters, no promise that the model will render readable text. The prompt defines the output, the subject, the mood, and the forbidden patterns.

Step 4 - generate a contact sheet

Do not judge one image at a time. Generate a set, then compare them as a contact sheet.

Use a simple scoring pass:

• 2 points: composition works
• 2 points: subject is clear
• 2 points: mood matches
• 2 points: no obvious visual artifacts
• 2 points: usable for the intended crop

Keep the top two. Reject the rest. This prevents you from spending five revisions on an image that had the wrong foundation.

Step 5 - revise one variable at a time

Bad revision:

Make it more realistic, more cinematic, less cluttered, warmer, sharper, and more premium.

Better revision:

Keep the same composition. Reduce desk clutter by half. Keep warm side light and muted paper palette.

One variable at a time is slower for the first two generations and much faster after that.

Step 6 - build a reusable shot list

Save the final pattern:

Shot: editorial desk comparison
Use: tutorial hero/card
Composition: solo creator, laptop left, notebook right, hands visible
Mood: warm side light, paper, ink, rust accent
Avoid: neon, glossy SaaS, fake readable text, extra fingers, brand logos
Revision rule: change only one variable per pass

Now the next article does not start from zero. It starts from a tested visual system.

Common mistakes

• Using too many references with no job assignment
• Asking the model to preserve text on screens or labels
• Changing prompt, aspect ratio, and references in the same revision
• Publishing an image that looks good at full size but fails in a cropped card
• Forgetting to check rights and resemblance before publishing

FAQ

How many reference images should I prepare?

Start with three: one for composition, one for mood, and one for material or texture.

Should all references be visually similar?

No. Similar references can trap the model. Use each image for a distinct job.

Can I use screenshots from other artists?

Use them for private study only. For publishable work, use your own photos, licensed references, or public-domain material.

What if the model copies the reference too closely?

Reduce the reference weight, describe the difference you want, and add a negative constraint such as "not a direct recreation."

Do I need a shot list for one image?

Yes if the image matters. A small shot list makes the first generation more intentional and makes revisions faster.