Midjourney Getting Started: Your First Image

Last updated：2026-08-29· 16 min read

🚀 Quick access

Midjourney V7 Domestic：Open entry↗
Text-to-image studio：Open mirror↗
Official Midjourney：www.midjourney.com ↗

Updated: 2026-08-29. Web / Discord entries, button labels, and parameter syntax follow midjourney.com and official docs that day. Midjourney has no official public API; third-party aggregators ≠ official service.

Introduction

This is a hands-on Midjourney beginner path: no long model reviews—just getting a first image you can hand off. You’ll prep account and acceptance criteria, choose Web vs Discord, write the first prompt, read the 2×2 grid, use Upscale / Vary, and add a few safe parameters. Beginners should finish the loop on whatever the UI shows as the current mainline model that day (community talk often centers the V8 line and the Niji track; confirm the on-screen default). For anime illustration, see Niji 7; for flagship depth, see the V8.2 guide.

What this guide solves

Finish account, acceptance criteria, and asset-rights prep before you generate
Understand Web vs Discord differences—and bind one learning session to one entry
Complete a first generate, read the grid, and use Upscale / Vary with intent
Use --ar and a few other basics safely, then self-check before delivery

Prep: ten minutes before you start

1. Pick an entry and write down identity

Choose one to start (you can try others later, but one learning pass binds to one entry):

Entry	Link	Notes
China-friendly (third-party)	Midjourney V7	Not Midjourney official; read privacy notes first
Mirror studio (third-party)	Text-to-image studio	Model list and queues may differ
Official	www.midjourney.com	Matches official account and docs

Write one line in notes: date + URL + on-screen model name. Network and account detail for China: China access guide.

2. Web or Discord?

Both can generate; the difference is mostly workflow habit:

Dimension	Web (official Create UI)	Discord (bot commands)
Who it fits	Visual UI, library management, fewer commands to memorize	Existing Discord habits, channel collaboration
Feel	Pick model, aspect, refs by click	Prompt + parameters in one message
Learning cost	Usually lower for beginners	Slash commands and button rows to learn
Collaboration	Mostly personal library	Channel visibility = privacy risk

Beginner recommendation: run Web until “prompt → grid → Upscale → Vary” clicks; migrate to Discord only if the team already lives there. If a third-party workbench is web-form only, follow that UI—but keep the same acceptance mindset.

3. Write a mini acceptance sheet

Fill the blanks before you open Create:

Use: ________ (e.g. 16:9 article hero)
Subject: ________ (one object/scene; state count)
Must have: ________ (whitespace placement, light, one style line)
Ban: ________ (text / watermarks / extra limbs / unauthorized brands)

Generating without an acceptance sheet means you cannot tell which step failed.

4. Assets and rights

Upload only references, logos, and pack shots you have rights to use
Faces, employee photos, client products need confirmed authorization
Do not ask to imitate a living artist’s unique style
Be cautious uploading sensitive commercial assets to third-party entries

5. Subscription and quota (no hard-coded prices)

Plan names, fast/relax modes, and remaining runs on official and third-party surfaces all follow the page that day. For learning, only track: how many usable runs left on this entry today. Do not paste old blog prices into budget sheets.

6. Three “fake problems” you’ll hit on the first failure

Beginners often misread process issues as “the model is broken”:

Conflicting prompts — “photo-real” and “anime line art” in one ask splits the grid. Cut down to one primary style.
Upscaling before composition passes — enlarging a bad crop only yields a sharper mistake.
Buttons don’t match the tutorial — Web, Discord, and third-party copy differ. Align to the acceptance sheet; don’t memorize U1/V1 letters as gospel.

Stick those three on a monitor sticky note—most anxiety disappears. Learning-path map: Guides hub.

Web & Discord: what to check on first open

Web checklist

Logged in; you can see Create / generate entry (labels follow the UI that day)
You can select or confirm the current model version
You can set aspect (or plan to append --ar at the end of the prompt)
You know where results land (job stream / library)

Discord checklist

Joined the server or bot channel Midjourney points you to that day
You know which slash command starts an imagine (often imagine-class; confirm that day)
After generate, a U / V (or similar) button row appears under the message
Public-channel generates are public—commercial drafts prefer private channels or Web

Family map: What is Midjourney.

Privacy and visibility: the beginner trap

Generates in public Discord channels may be seen, forwarded, or misused. Commercial drafts, unpublished products, client faces: prefer private Web space or private channels.
Third-party workbenches usually have their own history and log policies—read privacy notes; do not assume “same as official.”
After download, name files with project code and version (e.g. lamp-v1-u2.png) so you can find “which frame passed composition” a week later.

First image: step-by-step

We’ll walk a “ceramic desk lamp hero.” Swap the subject; keep the structure.

Step 1: choose aspect first

Set canvas before describing content. Common: article hero 16:9, square 1:1, vertical 9:16 / 2:3. Use the Web aspect control, or append --ar 16:9 (syntax per official docs that day).

Step 2: send the first prompt (no on-image text)

Use: 16:9 blog hero with ~30% whitespace on the right for a title.
Subject: one matte ceramic desk lamp, off-white shade, slightly left of center on a light wood desk.
Camera: eye level, mild side light, clean soft background.
Style: real product photography—no illustration, no sci-fi.
Ban: any on-image text, watermarks, borders, extra cables, warped shade.
--ar 16:9

If the UI already selected 16:9, omit a duplicate --ar to avoid override fights—follow how your entry behaves that day.

Step 3: read the grid

Midjourney typically returns a 2×2 grid. Do composition acceptance before you upscale:

Is the subject actually the lamp (not some other fixture)?
Is aspect close to target, with right-side whitespace?
Any obvious malformations (collapsed shade, floating object, broken perspective)?
Any banned elements (text, watermarks, mystery logos)?

Any fail → reroll the whole round or rewrite the prompt—do not enter fine edit. Pick the cell whose composition is correct (cells are usually 1–4 starting top-left).

Step 4: Upscale (single-cell refine)

Upscale the chosen cell (buttons may read U1–U4 or “Upscale”—follow the UI that day). Meaning: pull that cell into a standalone high-res result for download and further edit.

Important: Upscale cannot save a bad composition. Wrong crop → return to step 2.

Step 5: Vary for controlled iteration

Vary (or strong/subtle variants—names follow the UI) explores local differences once composition is right: slight light shifts, material tweaks, same camera with a color change.

Hard rules:

State what stays (camera, pose, palette, background)
Name one change only (if the entry allows an add-on prompt)
Re-check the acceptance sheet immediately

Keep the same composition, camera, lamp shape, and whitespace.
Only change the base material from matte ceramic to walnut wood.
Do not add text; do not change shade color.

For anime characters, switch to Niji and use an identity-sentence skeleton—see Niji 7.

Step 6: save and label

Download the usable frame; note something like v1-comp-pass-U2. All later edits constrain from that frame. When you need a reusable style library, read Styles & scenes.

Grid, Upscale, Vary: decision tree

Grid arrives
  ├─ No cell passes composition → rewrite prompt and reroll (do not blind Upscale)
  ├─ One cell passes → Upscale that cell
  │     ├─ Detail enough → delivery checklist
  │     └─ Needs tweak → Vary / ref-guided reroll (one change per round)
  └─ All cells “okay but wrong” → return to acceptance sheet; delete conflicting adjectives

Mental model for common actions (names follow the UI)

Action	What you’re doing	When
Regenerate	Same prompt, new grid	Bad luck; composition wiped
Upscale	Lock one cell as a standalone image	Composition already passed
Vary (subtle)	Small exploration	Light/detail only
Vary (strong)	Wider divergence	Composition OK, vibe wrong
Edit / inpaint	Region-specific change	Entry offers it and you already know how to accept

Third-party workbenches may merge or rename these actions—align to the acceptance sheet, not button letters.

Practice: same subject, three uses

Force yourself to change use, not adjective piles:

16:9 hero: right whitespace, no text.
1:1 hero: centered, solid background, contact shadow.
9:16 vertical: subject high; bottom reserved for later copy.

Each run: grid → composition acceptance → Upscale. You’ll feel immediately that aspect and whitespace are first-class constraints; style words are second. After that, the Prompt playbook reads much easier.

Core parameters: learn these three first

Full parameter thinking lives in the Prompt playbook. First image only needs:

Parameter	Intent (concept)	Beginner tip
`--ar`	Aspect ratio	Bind to use: hero 16:9, square 1:1
`--stylize` / `--s`	Stylization strength	Start default or mid-low; don’t max day one
`--weird`	Unusual / unconventional exploration	Skip while learning

Always: syntax, defaults, and ranges follow official docs that day. Old tutorial numbers may be dead.

Optional later (after the loop works): --sref style refs, character/object refs. Niji-track sref notes: Niji 7.

Safe parameter learning pace

Week 1: prompt structure + --ar only.
Week 2: once composition is stable, compare low vs high --stylize on the same prompt (one grid each).
Week 3: introduce refs when you need series vibe.

Do not day-one max weirdness, max stylize, and three reference images—you will not be able to attribute failures. Parameter meaning still follows official docs that day.

Iteration: one variable per round

Once you have a usable first frame:

State what stays
Name one change only
Re-check acceptance immediately; only then start the next round

How to roll back a bad edit

Return to the last correct Upscale and Vary again
Or open a new job: paste the acceptance sheet + a short description of the correct frame, optionally attach that frame as a reference
Avoid stacking ten contradictory edits on an already-drifting thread

References (beginner version)

Image 1: lamp silhouette and proportions (geometry must stay close).
Image 2: palette only (warm wood + off-white)—do not copy image 2 clutter.
Generate: 16:9, lamp on light wood desk, right whitespace; no text or watermarks.

Practice	Notes
Assign each image a job	Geometry / palette / scene mood—write them apart
Limit count	Beginners: one ref first
On conflict, write priority	“Shape from image 1; color from image 2”
Rights	Unlicensed images are not commercial delivery evidence

Products and logos: after generate, compare against brand source files item by item—“looks close enough” is not a pass. Style capture: Styles & scenes.

First-image timebox

Give yourself a 45-minute timebox. When it expires, stop and debrief—don’t all-nighter for luck:

Time	Do this
0–10 min	Fill acceptance sheet, pick entry, confirm model name
10–25 min	At most 3 grids; composition only
25–35 min	Upscale + at most 2 single-point Vary runs
35–45 min	Delivery checklist, download, notes

Still failing at timeout: the problem is usually the acceptance sheet or mutually exclusive prompts—not “ten more rolls.” Writing the timebox into personal norms builds reproducible skill faster than chasing one magic frame.

What to hand teammates

Solo practice can be screenshots; internal collaboration wants a delivery pack:

Final image file (with aspect info)
Full prompt text and parameter suffix
Entry + on-screen model name (that day)
Which cell won (e.g. “round-2 grid top-left → Upscale”)
Optionally one failed draft per failure mode (high teaching value)

Colleagues can then reproduce instead of asking which mystery button you clicked. Family and entry differences: What is Midjourney and China access.

Completion checklist (delivery)

Use and aspect are correct; crop / title safe zones are enough
Composition was accepted on the grid before Upscale
Subject type and count are correct; no obvious malformations or extras
Iteration did not introduce surprise text, watermarks, or wrong logos
If refs were used: consistency passes and rights are traceable
Checked at target display size (thumbnail + full)
Logged entry, model name, prompt summary, chosen cell
Know third-party ≠ official; sensitive data handled
Did not assume an official public API exists

Quick access

China-friendly (third-party): Midjourney V7
Mirror studio (third-party): Text-to-image studio
Official: www.midjourney.com

FAQ

Should I start on Web or Discord?

Learn the loop on Web first. If the team already lives in Discord, use private channels and watch public exposure. Third-party web UIs: follow that site’s controls.

Why does a long prompt make the grid messier?

Long prompts often conflict. Lock use, subject, aspect, and bans first; add one style layer at a time. Systematic writing: Prompt playbook.

Can I still edit after Upscale?

Usually you can keep Varying or regenerating with a reference, depending on the entry that day. Principle stays: start from a frame that is still correct; change one point.

Which model for anime?

Character / illustration track: evaluate Niji first. Photo-real products: mainline general models (e.g. V8.2 track). See Niji 7 and V8.2.

How much does it cost? Is it free?

Follow each entry and official pages that day; this guide does not hard-code unit prices or quotas.

Is there an official API?

No official public API. A third-party “API” is not Midjourney’s interface.

Official resources

midjourney.com
Official help and parameter docs (in-site entry that day)

Summary

Beginner success is not “one perfect poster on the first roll.” It is: acceptance sheet → pick entry → validate composition on the grid → Upscale to lock → Vary one point → tick the delivery checklist. Smooth that loop before you dive into parameter libraries and style cards. Version names, button labels, and prices always follow the official site or your entry that day—and do not assume an official public API exists.