How to Use Midjourney — V8 Complete Guide: Plans, Five-Layer Prompts, Parameters, and References

Midjourney is a paid AI image generation service that turns text (a prompt) into images. You can use it from the midjourney.com website or from Discord; as of September 2026 the default model is V8.2, and pricing runs across four plans from $10 to $120 a month. This guide walks through it in order, from your first image to choosing between the reference features.

On July 24, 2026, Midjourney switched its default model to V8.2. The V8 line started on March 17 on the alpha site only; V8.1 reached midjourney.com and Discord on April 30 and became the default on June 10. V8.1 made standard generation roughly 4–5× faster than earlier versions and added --hd for direct 2K (2048px) generation. V8.2 is an update focused on aesthetics, image quality, and Personalization, and on August 27 the rebuilt reference system, the Edit Model, started rolling out (all dates from the official Version page and the official Midjourney update posts). Midjourney itself announced V8 as "an entirely new model" that "may require totally new prompting styles," so V7 habits won't reproduce the same pictures.

Where you use it: both the web (midjourney.com) and Discord. Settings sync between the two, and the official docs say most features are offered on both platforms. That said, the Editor (redrawing a selected area), drag-and-drop reference images, and folder organization take fewer steps on the web. That bizarre initiation ritual — join a Discord server, paste commands at a bot — is no longer required.

My stance up front. For the average person starting out with image AI, Midjourney V8.2 is still the first pick. Stable Diffusion wins on freedom but its setup is heavy, ChatGPT/Gemini image generation is for inline throwaway use, and Adobe Firefly is commercially safe but weaker on artistic look. "Pretty images, from $10/month, with minimum steps, a fixed quota every month" — that demand still lands most cleanly on MJ. This article covers plan selection, prompt writing, parameters, references, and pitfalls, all based on the specs as of September 2026 (V8.2).

1. July 2026 — V8.2 Became the Default: What Changed in V8

Here is the V8 timeline on official dates. March 17: V8 opens on alpha.midjourney.com → April 14: V8.1 on alpha → April 30: V8.1 also on midjourney.com and Discord → June 10: V8.1 becomes the default (replacing V7) → July 24: V8.2 becomes the default → August 27: everyone can try the V8.2 Edit Model. V8.0 can no longer be used; its images simply remain in your gallery.

The biggest felt change in V8.1 is speed. The official docs put standard jobs at about 4–5× faster than earlier versions, and the announcement that made it the default says images appear in about 4 seconds in SD (1024px) and 12 seconds in HD (2048px). --hd is 2K generation with no upscaling step, costing about 1.3 minutes of GPU time versus about 0.8 minutes for SD. Quality and speed at the same time, finally.

The other big one is Personalization. Pick the images you like from a grid and the system builds a style profile tuned to you — then every future generation leans toward your taste from the start. Midjourney says V8.2 "understands" personal taste much better, especially on profiles with lots of ratings. "The same prompt produces different images for you and me" is the V8 norm. We've moved from the prompt era to the profile era, in my view.

2. What Is Midjourney — In Three Lines

Three-line breakdown:

Midjourney's strength comes down to one thing: "it looks great without much thought." While Stable Diffusion extracts top quality through stacked parameter tuning, LoRA selection, and ControlNet setups, MJ returns clearly above-industry-average results from a bare one-line prompt. You get results without the setup hellscape — that's what your $10/month buys.

Its weaknesses are equally clear. No public API (individual developers can't embed it); images are public by default, and only Pro and up can keep them private; NSFW and violent content are banned by the terms, and some inputs are blocked automatically. Many things that are trivial in SD are simply off the menu in MJ.

3. Pricing — A Four-Tier "GPU Hours" Model

MJ doesn't sell "X images per month." It sells Fast GPU time (how long you can generate in fast mode). The official rule of thumb is about 0.8 minutes per SD generation (a set of 4 images) and about 1.3 minutes for HD, so Basic's 200 minutes comes to roughly 250 SD runs — about 1,000 candidate images (official plan comparison and GPU Speed page, checked September 26, 2026).

My honest call: beginners should pick Standard ($30) without hesitation — it's the best value. Basic has no Relax mode (unlimited, slower image generation), so once Fast runs out you either buy more or wait for renewal — a constant stress if you iterate a lot. Standard gives you 15 Fast hours (roughly 1,100 SD runs) plus unlimited Relax, so you can stop counting. Move to Pro/Mega only when you want private generations, want to run video in Relax, or use it at a company with more than $1M in annual revenue.

4. Fast vs Relax — When to Use Which

This concept is specific to MJ and confuses newcomers. Let's lay it out.

5. Prompt Anatomy — The Five-Layer Structure That Works

Midjourney's own V8 announcement says the model shines with longer, more specific prompts (V8 Alpha post). Rather than listing whatever comes to mind, writing in the five-layer order below keeps you from leaving elements out.

A common beginner mistake: leading with feeling words like "a beautiful samurai". MJ interprets "beautiful" as a feature of the subject, drags the result toward "beauty," and the composition gets scattered. Replace feeling words with concrete or proper nouns (Hasselblad, golden hour, 85mm, etc.) and V8 reads you with much higher accuracy. That's why "people who know photography are usually better at MJ" is a real thing.

6. Seven Essential Parameters — --ar / --stylize / --chaos and Friends

Parameters are the --xxx flags you tack onto the end of a prompt. Seven you absolutely should know, with where they shine.

The two I personally use every single time: --ar and --stylize. The default 1:1 is awkward even for social posts, and stylize 100 is the middle of nothing. Make a habit of always appending --ar 16:9 --stylize 500 and your baseline output quality goes up. Reserve --chaos for exploration and --hd for final images. Note that --q (quality), which shows up in a lot of older guides, only works up to V7 and is not supported in V8.1/V8.2 (per the compatibility chart on the official Version page).

7. Reference Features — --sref / Edit Model / Moodboards / Personalization

This is where, in the V8 era, the gap between "beginners who fight with prose" and "advanced users who actually use the tool" shows up. Four reference-related features, broken down.

The most revolutionary one in my book is the "same subject from a reference image" capability that began with V7's --oref (Omni Reference). "Generate 100 images of the same character" used to require LoRA training and meant Stable Diffusion only. In V8.2 this moved into the Edit Model, which accepts up to four reference images at once (official Edit Model page). When features bleed between characters, Midjourney itself recommends combining the characters into a single reference image. That's one of SD's strongholds collapsing.

8. Three Common Pitfalls

Pitfall ①: Text rendering in V8 still favors short English

The V8 announcement calls text rendering "better than ever (when specified in quotes)," but the official docs themselves say it works best with the standard Latin alphabet and shorter words or phrases (Text Generation page). If a job needs reliable logo / poster / signage text — especially in non-Latin scripts such as Japanese — trying to finish it in MJ alone tends to fail. The practical play: use MJ for background + composition, then layer text in Photoshop / Figma — or hand text-first assets to Ideogram or ChatGPT's image generation. Don't expect "Logo with the text XYZ" to come out right from MJ alone.

Pitfall ②: Commercial use is allowed, but watch "public by default" and the $1M revenue rule

Under MJ's Terms of Service (effective May 27, 2026), you own the images you create to the fullest extent the law allows, and that ownership persists after you cancel. In return you grant MJ an irrevocable license to them, and images are public by default and can be remixed by others. Stealth mode, which keeps them private, is Pro and Mega only. On top of that, companies above $1M in annual revenue (and their employees) must be on Pro or Mega. Whether AI images are copyrightable at all varies by country, and MJ itself says it cannot give copyright guidance and suggests asking a lawyer. If you use MJ images in ad work, run it past legal.

Pitfall ③: No API = no app embedding

SD, GPT Image, and FLUX all have APIs and can be embedded into apps and workflows. MJ has no official public API — web and Discord only. Unofficial wrappers exist, but the Terms of Service prohibit using automated tools to access the service or generate images, so using one means accepting the risk of losing your account. If you need an API, pick another image AI. This has been MJ's stance since launch: "a tool that individual artists use by hand."

Summary

Midjourney is one of the few tools that has held its "king of image generation AI" status from 2022 onward. V8.1 raised speed and 2K output, V8.2 raised aesthetics and Personalization, the Edit Model raised references another notch, and many things that were "only Stable Diffusion" are now covered. For the demand of "pretty images, minimum steps, fixed quota every month," MJ is still the right answer. That said, if you need non-Latin text, an API, exact character matches, or cheap private generation, SD or other tools are still where you go. Treat it not as an all-purpose tool but as "a paid service specialized in artistry" — that's how you stick with MJ long-term.

FAQ

Related: Best 8 Image Generation AI Tools — Compared and Sorted by Use Case