The current Midjourney guide: prompt structure, every parameter that matters, personalization training, the 4Γ GPU trap, video mode, and a working loop.
Current default is V8.2 (July 24, 2026). Since V8.1 you get native 2K without a separate upscale step, standard generations that are 50% faster and 25% cheaper, and much better text rendering. The skill that matters most in 2026 isn't prompt trickery β it's training the personalization system to learn your taste, then iterating in tight loops.
The one trap to know before you start:
--sref, moodboards,--hdand--q 4run 4Γ slower and burn 4Γ the GPU time on V8. Budget accordingly.
V8 Alpha shipped March 17, 2026 on a completely rewritten codebase β not an incremental retrain. That's why the gains are structural rather than aesthetic: 5Γ faster generation, native 2K HD, and dramatically better text rendering.
Existing V7 styles and moodboards carried over. That was not guaranteed on a rewrite, and it's why the migration was painless for most users β accumulated style references didn't become worthless overnight.
Text rendering and prompt adherence were the two loudest complaints of the v6/v7 era. Both are substantially better in V8. Short labels, poster text and signage are now viable where they previously required a compositing step in another tool.
Prompt fundamentals, parameter grammar, and the iterate-don't-rewrite discipline. If you learned Midjourney on V7, you already know how to drive V8.
[subject] [action] [environment] [lighting] [style/medium] [parameters]
You don't need all six slots every time. You do need to know which one you're leaving empty and why β an unspecified environment means the model picks one, and it usually picks the most generic option available.
"Weathered fisherman mending nets at dawn" out-renders "portrait of resilience" every time. Abstract nouns give the model nothing to draw. If you can't photograph it, don't prompt it β convert the concept into the scene that would express it.
Golden hour, tungsten, overcast north light, rim-lit, single hard key. Naming the light moves results more than any style adjective, because lighting sets contrast, color temperature and mood simultaneously.
Stacked concepts dilute into mush. A prompt that wants to be cyberpunk and baroque and underwater produces something convincingly none of them. Pick the dominant idea; express the others through lighting and palette.
Vary the best of four rather than starting over. Rewriting throws away the seed that was already working. This is the most common beginner mistake and the easiest to fix.
| Parameter | Does | Use when |
|---|---|---|
--ar 16:9 | Aspect ratio | Always set it deliberately |
--s | Stylization strength | Low for realism, high for artistic license |
--c | Chaos / variety | Raise when the four results look samey |
--no | Negative terms | Standing hygiene: text, watermark |
--sref | Style reference | Series consistency |
--cref | Character reference | Same character across scenes |
--p | Personalization | Once you've trained it (below) |
--sref and --cref: The Series MakersThese two turn Midjourney from an image generator into an image-series generator: the same character in five scenes, the same brand palette across twenty assets. Both carried into V8 unchanged.
For brand work, --sref is the difference between "twenty nice images" and "a coherent campaign." It's worth spending a session building a reference set before you generate anything for delivery.
The practical gotcha nobody puts in the feature announcements: on V8, --sref, moodboards, --hd and --q 4 jobs run 4Γ slower and consume 4Γ the GPU time of a standard job.
Practical consequence: explore with plain prompts, apply style references only once the composition is locked. Running --sref during exploration burns your allowance four times as fast, for results you're going to discard anyway.
--s low (0β100) tracks your prompt literally and reads more photographic. --s high (500+) lets the model's aesthetic take over and reads more illustrative. Most disappointing results come from leaving it at default when the job needed one extreme or the other.
The V8 line's headline system learns your aesthetic from ratings and history. V8.2 reads taste noticeably better than earlier versions. Two users typing identical prompts now get different β and individually better β results.
--p once your profile has signal β using it too early amplifies noiseA larger model elsewhere doesn't know what you like. Personalization is the one advantage a competitor can't close by scaling: accumulated, per-user, non-transferable. Worth understanding as a user, because it means your investment in the platform compounds rather than resetting when someone ships a better base model.
Every downstream tool can fix a face, a hand, or a color cast. Nothing fixes a badly composed frame except regenerating it. Judging four candidates on composition β while ignoring the rendering flaws that draw your eye β is the highest-value habit in the whole workflow.
Cheaper, faster generations since V8.1 change the loop's math: generate more, curate harder. When each attempt cost real time, over-prompting to get it right first try was rational. It no longer is.
Animate any generated or uploaded still into 5β21 second clips with solid motion coherence. It's the fastest path from your image to motion.
It won't compete with Veo or Runway for cinematic work β no audio, no shot-to-shot continuity, no narrative sequencing. That's a positioning choice, not a gap someone forgot to fill.
Keep motion modest. Slow drifts, push-ins, atmospheric movement and fabric hold up far better than action. Ask for less motion than you want β under-driving produces usable output far more often than over-driving.
| Problem | Fix |
|---|---|
| Results look generic | More specific nouns; name the lighting; raise --c |
| Style drifts across a series | Lock --sref before generating any of the set |
| Text is garbled | Keep strings short β V8 handles labels, not paragraphs |
| Composition right, details wrong | Region edit, don't reroll |
| Character changes between images | --cref with a clean, front-facing reference |
| Everything looks like your last 50 images | Personalization overfit β prompt deliberately against your defaults |
| Burning GPU time unexpectedly | You're exploring with --sref on; turn it off until composition locks |
Rewrite the subject, not the modifiers. Nine times out of ten, a prompt that resists every parameter adjustment has a subject the model has no strong prior for. Change what the picture is of before you change how it looks.
Plans run $10β$120/month across four tiers. The variable that matters isn't the feature list β it's GPU time, which is what you actually consume.
--sref habitually--sref is on constantly and video is part of the outputEstimate consumption as standard jobs Γ 1, reference/HD jobs Γ 4. Most people who feel they've outgrown a tier have actually just started using style references and never adjusted the mental model.
For V7-era fundamentals and what carried over, see the V7 tutorial. For how Midjourney stacks against GPT Image 2 and Flux in 2026, see the image-tools comparison. For version-by-version changes, the live updates tracker.
Sources: Midjourney official version docs Β· WaveSpeed β What is Midjourney V8 Β· AIxploria β pricing and --cref/--sref prompt guide Β· Blake Crosley β Midjourney 8.1 guide Β· TechJack β prompt structure and parameters Β· Beginners in AI β V8 explained
Last updated: July 30, 2026
A practical six-stage launch guide for AI tool founders, with 30 copyable actions and clear records for owners, dependencies, evidence, and results.
A practical guide to evidence, ten checks, editorial scoring, commercial disclosure, and the limits of an AI tool assessment.
How Toolso structures multilingual URLs, canonical tags, hreflang, sitemaps, and translation fallbacks, with checks for directory builders.
How to produce AI video in 2026: picking a model per shot, prompting for motion, keyframe workflows, real cost math, artifact fixes and Sora migration.
The 2026 automation guide: how Zapier, Make and n8n differ on pricing model, AI agents and self-hosting β plus when a workspace agent beats all three.
Prompt engineering for 2026: why leaner prompts score 10β15% higher, what reasoning modes made obsolete, the five patterns that survived, and agent specs.
The 2026 playbook for coding with Claude: lean CLAUDE.md, plan-first workflows, subagents, parallel worktrees, hooks, and verification loops that work.
A working guide to Copilot in 2026: Agent Mode setup, auto-approve policy, agentic review, cost discipline, and the habits that make agents net-positive.