Midjourney Image Reference in Practice: Controlled Variants from Reference Images

This tutorial walks you through a complete Midjourney image-to-image / image reference exercise: from choosing a reference and writing prompts to SD validation and Run as HD delivery. The goal is not to “get lucky with something that looks like the reference,” but to build a repeatable workflow for controlled variants.

Best if you already know basic text-to-image. If you have not completed your first generation yet, start with Getting Started with Midjourney: Your First Complete Generation Guide. For principles and extended scenarios, read Midjourney Image-to-Image & Image Reference: From One Reference Image to Controlled Delivery.

What You Will Be Able to Do

After completing this lesson, you should be able to:

1. Decide when to use image reference and when to stay with pure text-to-image
2. Pick a reference image that “does only one job”
3. Use text to spell out clearly what to change and what to keep
4. Deliver controlled variants under Midjourney V8.1 with an SD→HD workflow

Estimated time: 20–30 minutes.

Before You Start

• A Midjourney workspace account (enter via this site’s “Try Midjourney” entry point)
• 1 clear reference image (recommended: white-background or clean-background product shot / medium-close character headshot)
• A concrete goal, for example: “Same product, moved to a sunny breakfast table scene”
• Confirm the model is V8.1

For practice, product still lifes work well: structure is clear, so you can tell immediately whether the reference is doing its job.

Exercise Overview (Read the Map First)

Step	What to Do	Success Criteria
1	Define goal and no-change items	One sentence that states “keep / change”
2	Choose and inspect the reference	Subject complete, clear, not overly complex
3	Text-only direction check (optional but recommended)	Subject and framing roughly correct in SD
4	Image reference + lean prompt	Structure like reference; scene changes via text
5	Change only one variable for comparison	You can point to which prompt line or setting worked
6	Small steps with Vary / Remix	Better detail without drifting off target
7	Run as HD delivery and log notes	~2K deliverable + reusable notes

Step 1: Write “Keep / Change” in One Sentence

Before opening the workspace, write two lines:

• Must keep: e.g. “bottle shape, label layout, brand colors”
• Must change: e.g. “scene → oak breakfast table + morning window light”

If these two lines are unclear, the reference and the text will fight each other later.

Step 2: Pick a Reference That “Does One Job”

What you want to lock	Better reference	Not recommended
Product shape / placement	Clean still life, subject complete	Multi-product collage, heavy-filter cover
Character vibe	Medium-close with readable face lighting	Full-body wide shot with blurry face
Composition skeleton	Finished shot with clear subject relationships	Overloaded poster layout

Rule: one reference image = one primary job. For unified look across a series, use SREF / Moodboards separately (see the tutorial Series Visuals with Style Reference: Brand Consistency in Midjourney); do not expect one image to lock composition, style, and character at once.

Checklist:

• Key structure of the subject is not cropped away
• Resolution is sufficient; no obvious compression blur
• Background does not steal focus (or you accept that it may transfer too)

Step 3 (Recommended): Confirm Direction with Text-to-Image First

Many people attach a reference right away, then cannot tell whether failure came from composition or style. Run 1–2 pure-text rounds in SD first:

ceramic skincare bottle on a sunlit oak breakfast table, soft morning window light, lifestyle product photography --stylize 100

Check whether framing and scene are close to your “must change.” If direction is wrong, fix the text first; if direction is right, move on to image reference.

Step 4: Add Image Reference and Write “What Changes”

1. In the workspace, add your image reference (image prompt) for this generation
2. Focus the prompt on changes; do not repeat shape details already obvious in the reference
3. Stay in SD during exploration

Example you can adapt directly (product, new scene):

same product silhouette and label layout, place on a sunlit oak breakfast table, soft morning window light, lifestyle product photography --stylize 100

Writing tips:

• same … phrases make the structure you want to keep explicit
• The second half describes only the new scene, light, and lens feel
• Avoid style words that fight the reference (e.g. reference is clean commercial photography, but you write “cyber-neon thick impasto”)

For commercial realism, try adding --raw so results stick closer to the reference and literal prompt. Parameter meanings: see Midjourney V8.1 Parameter Cheatsheet: –hd –raw –chaos –stylize.

Step 5: Change Only One Variable at a Time (Build Control)

Keep the reference fixed and run three experiments, changing only one thing each time:

1. Light only: morning window light → soft overcast daylight
2. Scene only: breakfast table → bathroom counter
3. Toggle --raw only: compare realism obedience

View the three results side by side. You will quickly see: image reference handles the “skeleton,” text handles the “differences.”

Step 6: Refine with Vary / Remix, Not a Full Restart

When you have a candidate that is “structure right, details slightly off”:

1. Use Vary Subtle to find a better composition in nearby space
2. Use Remix when you need to change the description while keeping structural wins
3. If the look still drifts, add SREF instead of piling on more adjectives

For multi-image series: image reference locks the subject; style anchor locks the vibe—do not mix those roles.

Step 7: Run as HD Delivery and Final Check

Once direction is locked:

1. Apply Run as HD or --hd to the winning result (~ 2K HD)
2. Check: do label / shape still match “must keep”?
3. Check: do scene and light match “must change”?
4. Save: reference notes, final prompt, key parameters, deliverable filename

Note: V8.1 / V8.2 do not support the legacy quality flag --q. Deliver with an SD→HD strategy; do not copy V7’s --q 2 / --q 4 blindly. Resolution strategy: see Midjourney HD vs SD: Save Time and Cost with a Resolution Strategy.

Class Exercise: Pick One of Three (Finish Any One)

Exercise A: E-commerce Product, New Scene

Reference: white-background product shot → generate one each for kitchen / bathroom / travel scene (same reference, change scene text only).

Exercise B: Character Headshot, New Background

Reference: medium-close headshot → generate café window and neon night-street versions while keeping facial vibe.

Exercise C: Packaging Material Tweak

Reference: front of pack → keep shape, change material from matte to glass look (change only the material sentence in text).

Every exercise follows: SD validation → single-variable comparison → HD delivery → log prompts.

FAQ

Why does the result still not look like the reference?

Common causes: subject cropped in the reference; text conflicts with reference style; too many variables changed at once; you needed style reference but used only image reference. Go back to your “keep / change” two lines and redo Steps 4 and 5.

Is image-to-image “editing the original image”?

No. Midjourney still regenerates from scratch: image reference adds constraints so results are controllably similar, not pixel-level retouching.

How does this work with SREF / Moodboards?

• Structure / shape: image reference
• Brushwork / palette / series vibe: SREF / Moodboards
• Content narrative: text-to-image prompt

Full stacking order: see Midjourney Style Reference Complete Guide: How SREF and Moodboards Keep Series Visuals Cohesive.

What to learn next?

• Prompt skeleton: Midjourney Prompt Tips: A Structure Built for V8.1
• Templates to adapt: Midjourney Prompt Case Library: 10 Ready-to-Adapt V8.1 Scene Templates
• End-to-end workflow: Midjourney AI Art from Start to Finish: The Complete 2026 Workflow (V8.1)

Once your reference is ready, enter the workspace from Try Midjourney in the header and run through all seven steps for your first set of controlled variants.

Summary

The core of Midjourney image reference in practice is constraint, not luck: image reference locks structure, text locks differences, SD buys information, Run as HD delivers 2K HD. Finish one exercise through all seven steps and you have a reusable controlled-generation path—closer to real creative and commercial delivery than occasionally landing a “miracle shot.”