How to Use Midjourney AI: A Practical Walkthrough

Getting Started: Where Midjourney Lives Now (and What It Costs)

If you searched for Midjourney tutorials a year ago, you'd have found Discord commands and bot invites. That era is over: the web app opened to all subscribers on August 23, 2024, and it's now the primary interface for everything from generation to editing and moodboards, with Discord relegated to legacy community status.

This matters practically, because several of Midjourney's newer capabilities — video generation, personalization, and advanced editor tools — are web-only. If you're following an older tutorial that walks you through /imagine commands, you're learning a workflow the platform itself has moved past. Start at midjourney.com, sign in with your subscription, and work from there.

The next reality check is cost. Midjourney discontinued its free trial on March 28, 2023, so there's no free tier — you're committing at least $10/month for the Basic plan. The full ladder looks like this:

• Basic ($10/mo) — 3.3 hours of Fast GPU time and roughly 200 image jobs
• Standard ($30/mo) — 15 Fast hours plus unlimited Relax mode
• Pro ($60/mo) — 30 Fast hours, Stealth Mode, and higher concurrency
• Mega ($120/mo) — 60 Fast hours for heavy production use

Annual billing saves about 20% if you're confident you'll stick around, and extra GPU time costs $4/hour if you burn through your allotment.

Two caveats deserve special attention before you pick a plan. First, businesses earning over $1 million per year must subscribe to Pro or Mega — this isn't optional, it's written into Midjourney's terms, so companies evaluating it for branded content need to budget accordingly. Second, your images are public by default. Stealth Mode, which keeps generations private, sits behind the Pro plan's $60 paywall — a departure from industry norms that many reviewers, including the team at Memeburn, find odd. If you're generating proprietary concepts, product mockups, or client work, factor that privacy cost in from the start.

One more consideration: there's still no public API, so integration with your existing toolchain means manual work or third-party wrappers. And since AI-image copyright law remains unsettled, check the rules in your jurisdiction before putting generated images into commercial assets.

Getting started, then, is a two-step decision: confirm the web app fits your workflow, and pick the plan that matches both your volume and your legal situation. For solo creators testing the waters, Basic at $10/month is a reasonable entry point. For anyone generating at scale or under a business umbrella, the Pro-tier requirements and privacy trade-offs are the real gatekeepers — sort those out before you generate your first image.

The Core Generation Loop: Prompt, Four Images, Vary or Upscale

Every image you've ever admired from Midjourney came out of one deceptively simple loop: describe what you want, review four candidates, then either remix or refine. Master that cycle and the rest of the tool becomes second nature.

You type a prompt, and Midjourney returns four image variations in roughly 50 seconds on a basic plan, according to CMSWire's hands-on testing. Once the grid appears, you face the core decision: reroll for four fresh takes, generate variations of the one image that caught your eye, or upscale a favorite — and upscaling takes about a second, so the slow part is always generation, never refinement.

How much you can queue up depends on your plan. Basic and Standard subscribers get 3 concurrent jobs with a 10-job queue, while Pro and Mega users can run 12 concurrent fast jobs plus 3 relaxed ones. That concurrency headroom matters more than it sounds: heavy iteration sessions stall quickly when you're waiting on three slots.

When you do upscale, you have two options introduced with V6: Subtle and Creative 2× upscalers. Subtle preserves the original's character at double resolution; Creative takes liberties and adds detail as it enlarges, as detailed in the Midlibrary model review. Either way, you top out at a maximum resolution of 2048×2048 px — larger than DALL-E 3's 1536×1024 output, per G2's comparison testing.

A few practical habits make the loop far more productive:

• Batch your prompts — with only 3 concurrent jobs on lower tiers, plan prompts while generations run.
• Vary first, upscale second — variations cost the same generation time but explore more aesthetic territory.
• Use Subtle upscale when a composition is already right; use Creative when the image needs more texture.
• Expect to re-roll often — Midjourney adds its own artistic spin rather than rendering prompts literally.

That last point deserves emphasis. Multiple independent tests found Midjourney "likes to get creative" and interprets prompts loosely, which is precisely why the four-image grid exists — you're auditioning interpretations, not commissioning exact renders. At SkillzBench, we think about verification the same way: tools earn trust through repeated, structured testing rather than a single impressive result. Budget your GPU time accordingly, and treat each 50-second generation as one round of an ongoing conversation with the model.

The Parameters That Actually Matter: --style raw, --stylize, --ar, and --chaos

Midjourney has a personality — it "likes to get creative" and adds its own artistic spin to whatever you type, as G2's hands-on comparison found. The right parameters let you decide how much of that personality shows up in your four generated images.

--style raw is the single most useful switch when you want literal results. Introduced with V6, it reduces Midjourney's default artistic flair so the model follows your prompt more closely, according to Andrei Kovalev's in-depth V6 review. Use it for product shots, reference-style imagery, or anything where accuracy beats atmosphere.

--stylize controls the opposite direction: how much aesthetic interpretation Midjourney applies. Higher stylize values lean into the "lushness" the model is famous for; lower values trade beauty for prompt fidelity. This trade-off matters — Jakob Nielsen, who generated roughly 20,000 AI images in 2025, concluded that "lushness buys attention; prompt adherence buys meaning," and meaning wins when an image needs to teach something.

--ar sets your aspect ratio, which is essential for matching output to its destination — wide banners, vertical social posts, square thumbnails. --chaos increases variation across the four images in a grid, useful when you want to explore a concept broadly rather than refine one idea.

Kovalev's testing surfaced two practical fixes worth memorizing:

• Composition: Midjourney favors asymmetry and off-center subjects. Adding "central composition" or "symmetrical" to your prompt restores centered framing.
• Text in images: put your words in quotation marks, lower the stylize value, use short words, and repeat the text in the prompt.
• Expect re-rolls: Kovalev's verdict on text rendering — "The key word here is perseverance" — assumes countless attempts before a near-perfect result.

That last point deserves emphasis. When Memeburn tested a sign reading "DREAM BIG," one of four images rendered "DREAM DREAM BIG" and the rest came out blurry. Budget your GPU time accordingly — a basic plan includes only about 3.3 hours of fast GPU time per month, so text-heavy projects can burn through it quickly.

The same test-first mindset applies beyond Midjourney. At SkillzBench, we're building benchmarking infrastructure that measures whether AI tools and agent skills actually improve results — the same discipline of hands-on verification that separates useful parameters from marketing claims.

Knowing Midjourney's Limits: Prompt Adherence, Text, and When to Use Something Else

Midjourney will happily draw you a gorgeous cat that also happens to be a unicorn — even when you asked for a cat riding a unicorn. That's the tool in miniature: stunning output, imperfect obedience. Knowing where it fails saves you hours of re-rolls.

The failure modes are well documented. In G2's fantasy test — a cat riding a unicorn through neon-lit Tokyo — Midjourney produced unicorn-cat hybrids and hornless horses, showing weaker prompt fidelity than DALL·E 3. CMSWire's reviewer asked for two moons and got the wrong count; only refining the prompt wording improved the result. Text is worse still: when Memeburn requested a sign reading "DREAM BIG," one of four images rendered "DREAM DREAM BIG" and the rest came out blurry.

Jakob Nielsen, who generated roughly 20,000 AI images in 2025, distills the tradeoff into a single line: "lushness buys attention; prompt adherence buys meaning. And meaning wins when the image is supposed to teach." His own usage tells the story — Midjourney produced only 2 of his top 12 images in 2025, down from 7 of 10 winners the year before.

That insight translates into a simple practical rule:

• Use Midjourney for mood, storytelling, and emotional impact — hero images, concept art, and atmospheric visuals.
• Lean on it for stylized, cinematic photorealism; G2's testing found its lighting and skin texture genuinely convincing.
• Switch to a literal tool like DALL·E 3 when you need precision: accurate multi-concept scenes, legible text, or instructional graphics that must teach.
• If you must render text in Midjourney, expect perseverance — countless re-rolls and prompt updates, per Midlibrary's reviewer.

One more caveat before you ship anything: AI-image copyright law remains unsettled, with no clear legal stance on ownership of generated content. Laws are changing constantly, so verify the rules in your country and state before using images in branded content — and remember that businesses earning over $1M annually must be on Midjourney's Pro or Mega plans.

The same discipline applies across the AI stack, not just images. At SkillzBench, we're building exactly this kind of evidence-first evaluation for AI agent skills — benchmarking each skill in a sandbox, security-scanning every version, and publishing the measured lift so you know what actually works before you rely on it. Whether you're choosing an image model or an agent skill, the principle holds: match the tool to the job, and trust test results over marketing. If you want vetted, benchmarked agent skills as that ecosystem matures, the SkillzBench waitlist is open.

From Images to Workflow: Making Midjourney Part of a Verified AI Stack

Generating a striking image is easy; making image generation a dependable part of your workflow is not. The teams that get consistent value from Midjourney treat it as one component in a stack — with a repeatable process for prompting, refining, and verifying output before it ever reaches a customer.

Start with a prompt-refine loop you can repeat. Midjourney's core rhythm — prompt, four images in roughly 50 seconds, then variations or an upscale in about a second — is fast enough to iterate aggressively, according to hands-on testing. Log what you changed and what improved; CMSWire's reviewers found that simply refining prompt wording corrected misinterpreted details like moon counts. Keep --style raw for literal subjects, --stylize for aesthetic range, and the Subtle or Creative 2× upscalers for final output.

Before you commit to a plan, run a short selection check:

• Budget: pricing starts at $10/month with no free tier, and annual billing saves 20%.
• Volume: Basic includes 200 image jobs and 3.3 hours of Fast GPU time; Standard adds unlimited Relax mode.
• Privacy: images are public by default unless you're on Pro or Mega with Stealth Mode.
• Business use: companies earning over $1M annually must use Pro ($60/mo) or Mega ($120/mo) plans.

Then do your legal homework before commercial use. There is still no clear legal stance on copyright ownership of AI-generated content, and legal experts advise checking the laws in your specific country and state before using AI images in branded content. Litigation is already live — Evox Productions sued Midjourney over copyrighted car images — and the digital artist community's objections to text-to-image tools remain unresolved.

Finally, know when to route work elsewhere. Jakob Nielsen, who generated roughly 20,000 AI images in 2025, used Midjourney for only 2 of his top 12 images — down from 7 of 10 the year before — concluding that "lushness buys attention; prompt adherence buys meaning." Use Midjourney for mood and aesthetics; use more literal tools for instructional or text-heavy graphics.

That same verify-before-you-trust discipline applies to the rest of your AI stack. SkillzBench, AIQ Labs' independent testing lab for AI agent skills, is building exactly this for the instruction packages that power Claude Code, Cursor, Gemini CLI and other agents: Discover to crawl and dedupe skills from GitHub, Verify to pin every version to a commit SHA and run license checks and security scans, and Benchmark to test each skill in a sandbox — with and without it — and publish the measured Lift. Published research found roughly 26% of skills carry at least one vulnerability, so proof matters. If you're assembling an AI toolkit where every tool is proven rather than merely promising, join the SkillzBench waitlist at skillzbench.com — skills, proven.