Midjourney
Curriculum — v7 / v8.1 era

Platform Ground Truth

Before any exercise, internalize the current state. Most outdated guides will burn your time.

Models live right now

• V7 — current default. Released April 3, 2025; became default June 17, 2025. Introduced Draft Mode and Omni Reference. Best for photoreal, complex natural-language prompts, text rendering, and character consistency via --oref.
• V8.1 — released on midjourney.com April 30, 2026. Fastest model so far. Standard jobs render 4–5x faster than earlier versions. HD images at 2048px without a separate upscaler. Personalization Profile must be unlocked to use it. Sharpness improvements especially with SREFs and moodboards.
• V8.0 Alpha — launched March 17, 2026 on alpha.midjourney.com. Only compatible with Fast mode. Skip unless you have a specific reason.
• Niji 7 — anime/manga specialization. Use only when the brief calls for that register.

The Parameter Surface

Verified current. Memorize this table; everything downstream assumes it.

Things to drop from your vocabulary immediately

• --cref and --cw. V6-era. When using V7, use Omni Reference instead.
• “Masterpiece, 8K, trending on artstation, hyper-detailed, award-winning.” V4/V5 cargo cult. V7/V8.1 read these as noise.
• --q 2 reflexively — not justified in V7/V8 unless you have a specific reason, and not supported on V8.1.
• Stacking 5+ artist names. The model averages references into mush.
• --chaos 100 --weird 3000 --s 1000 together. One knob at a time.

Official parameter docs: docs.midjourney.com

Day Zero — Rate the 200 Pairs Deliberately

This is the highest-leverage hour you will spend in Midjourney, and almost nobody treats it seriously.

The Mental Frame

You’re not picking which image is “better.” You’re not picking which one you’d hang on a wall. You’re training a model that will live behind every prompt you write for the next year.

Reframe the question from “which do I like more?” to “which one is closer to the work I actually want to make?”

Those are different questions. A glossy, well-lit, beautifully composed fantasy portrait might be “better” by any conventional metric. But if you click it, you’ve just told Midjourney that’s the direction. Every horror prompt you write afterward will have that glossy fantasy pull working against you.

The Default You’re Fighting Against

Midjourney’s untrained aesthetic — the thing your profile is supposed to override — biases hard toward:

• High saturation, warm color grading, golden-hour everything
• Symmetrical composition, centered subjects
• Clean skin, clean surfaces, no texture noise
• Romantic-fantasy register (think: digital concept art, ArtStation top page)
• “Epic” lighting — volumetric rays, lens flares, rim lights
• Shallow depth of field as default, regardless of subject
• Conventional beauty in faces
• Resolved, finished, polished

Your register is the opposite of almost every item on that list. So your rating job is to systematically pick against this default.

The Rules of Thumb (apply in this order)

When a pair comes up, ask these questions in order. Stop at the first one that resolves the pick.

1. Which one has more texture/grain/imperfection? Pick the grittier one. Film grain, halation, gate weave, dust, scratches, video noise, scanlines, compression artifacts — all of these are good. If one image looks cleaner and one looks rougher, pick rougher every time.
2. Which lighting is more motivated and less pretty? A single hard practical light (desk lamp, fluorescent tube, TV glow, car headlight, streetlight) beats a beautiful golden-hour wash every time. Hard shadows beat soft. Underexposed beats correctly exposed. Color contamination (tungsten in one corner, fluorescent in another) beats clean white balance.
3. Which one is more uncomfortable to look at? Off-center composition, dead-center deadpan staging, awkward eye contact, weird negative space, a subject too close to the edge of frame, a horizon line in the wrong place — pick the one that makes your eye uneasy. Conventionally well-composed is the enemy.
4. Which palette is sicklier, more contaminated, more limited? Sodium yellow, fluorescent green, dental-office cyan, dried-blood maroon, nicotine-stain beige, parking-lot orange. Avoid: teal-and-orange, warm gold, lush jewel tones, anything that looks like an Instagram filter.
5. Which one has more deadpan or dread in it? A neutral face beats a smiling face. An empty room beats a busy room. A static composition beats a dynamic action shot. Stuart Gordon and the Safdies live in flat affect; train for it.
6. Which one looks less like AI? The hardest and most important. AI tells: hyper-resolved everything, no aliasing, perfect facial symmetry, hands that look correct but lifeless, lighting that comes from everywhere, surfaces that look 3D-rendered. Pick the image that looks more like it was captured by a camera (with all the lens flaws, focus misses, and grain that implies) than rendered.
7. Tiebreaker: which one is weirder? If neither image violates the defaults above, pick the stranger one. Your register rewards weird.

Specific Pair Types You’ll Encounter

Portrait vs. Portrait

Pick the one with harder lighting, more shadow on the face, less symmetry, less conventionally attractive features, more skin texture, more lived-in expression. Reject magazine-cover faces. Reject the one with the lens flare.

Landscape vs. Landscape

Pick the one with worse weather, less golden-hour, more haze, more emptiness. A flat overcast field beats a dramatic mountain at sunset. Pick the one that feels like nothing’s happening.

Two Abstract / Textural Images

Pick the one closer to film grain, paper, fabric, decay, mold, rust, fluorescent buzz. Avoid the one that looks like polished digital art.

Interior vs. Interior

Pick the one that looks more like a real place someone has been miserable in. Wood paneling, drop ceilings, cheap carpet, fluorescent tubes, popcorn ceilings. Reject the one that looks like an architectural rendering or a hotel lobby.

Animal / Creature Pair

Pick the more anatomically wrong one. Pick the one closer to body horror. Pick the one that looks like it has weight, not the one that looks airbrushed.

Genre-coded Pairs (sci-fi, fantasy, etc.)

This is where most people accidentally train themselves into a corner. If you see two “cool sci-fi soldier” images, pick the one that looks more like Possession-era European weird than Halo cover art. If you see two “fantasy castle” images, pick the one that looks more like a Tarkovsky frame than a Disney background.

When Both Are Bad

Pick the one that’s bad in a more interesting way. A failed render that looks like a glitch is better than a failed render that looks like a generic stock illustration.

When Both Are Genuinely Good and Both Fit Your Register

Pick the one with more restraint. Less is more.

What to Watch Out For

After You’re Done — Sanity Check

Generate this exact prompt to test your profile:

A man sitting alone in a wood-paneled motel room at 2am, single 
tungsten lamp on the nightstand, television off, curtains drawn, 
shot on 16mm Kodak Vision3 500T pushed one stop, handheld, 
medium wide, deadpan composition --ar 2.39:1 --s 100 --p

If the output looks grainy, underexposed, deadpan, slightly sickly in palette, with hard motivated lighting and visible film texture — your profile is calibrated correctly.

If the output is warm, glowing, beautifully lit, with a handsome subject and an “epic” feel — you rated too conventionally. Rate another 100 pairs being more aggressive about picking the rougher option.

The Methodology

Mastering Midjourney is mastering a feedback loop. A first prompt is a hypothesis; the four-up grid is the data; iteration is the work.

Phase order: Brief → Moodboard (Draft Mode, chaos 40–60) → Direction lock (chaos 10–20) → Style lock (--sref documented) → Character/object lock (--oref documented) → Hero generation (chaos 0–5, full quality, seed noted) → Document.

The single biggest discipline upgrade: every locked look gets recorded — prompt template, sref code or URL, sw value, stylize value, example output. Treat this like a LUT library. If you can’t reproduce a look, you haven’t made it.

Prompt Anatomy

V7-era prompts are paragraphs, not tag lists. Keep text prompts simple — avoid adding style words that conflict with your reference image’s look. Focus on content, not instructions: use your text prompt to describe what you want to see, not how Midjourney should modify the reference.

The Six Functional Slots

1. Subject and action. Who, what, doing what. Specific verbs.
2. Environment. Where, what’s around, what the world is made of.
3. Composition and camera. Shot type, lens, angle, framing.
4. Lighting. The single highest-leverage variable. Source, direction, quality, color temperature.
5. Style/medium reference. One named cinematographer or photographer, or an SREF. One or two anchors maximum.
6. Parameters. Always --ar. Then stylize, chaos, sref/sw, oref/ow as needed.

Skeleton

[Subject doing something specific] in [environment with material 
detail]. [Shot type, lens, angle]. [Lighting source, direction, 
quality, color]. Shot on [film stock or sensor], [grain/texture]. 
In the visual register of [one or two references]. 
--ar 2.39:1 --s 150 --c 10

The Contrast That Should Be Welded Into Your Reflexes

V7/V8.1 respond to physical detail, optical detail, material detail. Your filmmaking vocabulary is the unfair advantage.

Grammar Through Frames

Goal: internalize aspect ratio, stylize, chaos, seed, draft mode, and the difference between abstract and physically specific prompts. Do this while shot-matching, not before.

• Unlock your V7 Global Personalization Profile (use the methodology in § 03 above). This is the gate to Draft Mode and to V8.1. Do not skip.
• Read the official Version page top to bottom: docs.midjourney.com / Version
• Read the Parameters overview: docs.midjourney.com / Documentation

Pick one film with strong visual identity. For your register: The Thing (1982), Possession (1981), Hereditary, Bound, Twin Peaks: Fire Walk With Me, Uncut Gems, Suspiria (1977), Mandy, Beyond the Black Rainbow. Pick one.

Pull five stills. For each still, work the loop:

1. Write a first-draft prompt in Draft Mode at --c 20 --ar [match the still]
2. Diagnose the gap between grid and source. Is it lighting? Lens? Palette? Grain? Composition?
3. Change one variable. Re-generate.
4. Repeat until match or until you understand why it can’t match (sometimes it’s an --oref situation you’ll handle in Week 3).
5. Note the prompt evolution in a journal. This journal is non-negotiable.

By end of Week 1: ~5 stills taken to genuine match, internalized --ar, --stylize, --chaos, --seed, and the difference between underspecified and physically specific.

Week 1 Resources

• Draft Mode official doc: docs.midjourney.com / Draft & Conversational Modes
• Parameter behavior reference (independently maintained, more current than most): blakecrosley.com / Midjourney guides
• Future Tech Pilot for parameter intuition (still active; verify videos are V7-era vs. older): youtube.com / @FutureTechPilot

Style References & Lookdev

Goal: build your first reusable “look.” Move from one-off generations to a system.

• Official doc: docs.midjourney.com / Style Reference
• Take the two strongest results from Week 1. Use each as a style URL via --sref [image URL]. Generate a different subject (different person, different environment) in that locked look.
• Run a --sw sweep: same prompt and sref, with --sw 50, --sw 150, --sw 300, --sw 500. Range 0–1000, default 100. Observe what changes.

• Each numerical code produces a unique visual style: --sref [code]. Use --sref random for a random style code.
• Run 10–15 --sref random generations on the same subject prompt. When you hit one that resonates, note the code in your journal.
• The Style Reference feature was updated for V7; old style codes may not produce the same styles. To use old codes, add --sv 4 to your prompt. Worth knowing if you find a 2024-era code in someone’s library.
• Deep-dive on sref codes: midlibrary.io / deep dive into sref codes

• Lock down a complete recipe for ONE register you’ll use repeatedly. For you: maybe “16mm horror nighttime exterior” or “Reagan-era VHS public access.”
• Recipe must specify: subject template, environment template, sref code or URL, --sw, --stylize, --ar, an example output. Save the four-up grid.
• Generate three different subjects through the same recipe. They should look like one cinematographer shot all three. If they don’t — sref too weak, prompt too dominant, or stylize too high — diagnose and fix.

You can stack: --sref URL_A URL_B or --sref code_A::2 code_B::1 to weight them. Try blending two references.

By end of Week 2: one fully documented “look” recipe. This is your first LUT.

Week 2 Resources

• Andrei Kovalev’s Midlibrary (best independent sref catalog and analysis): midlibrary.io
• Tatiana Tsiguleva’s Substack (working designer publishing live V7/V8.1 experiments): ciguleva.substack.com

Character/Object Consistency & Sequence

Goal: build a coherent visual sequence — same character, same world, multiple shots.

• Official doc: docs.midjourney.com / Character Reference
• On the website: drag and drop your reference image directly onto the imagine bar where it says “Omni Reference.” On Discord: type --oref followed by the URL.
• Omni Reference works very well with images from outside Midjourney — higher consistency than V6’s cref, even works for putting yourself into scenes.

Default --ow is 100. Range 0–1000. Low values (25–100): minimal influence, more AI creativity. Higher values: stronger lock.

• Generate a character portrait you like. Use it as your --oref.
• Generate the same character in 6–8 different scenes: close-up, wide, dialogue moment, action moment, exterior day, interior night, different costume, looking at camera vs. away.
• Sweep --ow: 50, 150, 300, 500. Find where consistency holds without freezing the character into one expression.
• Document the working --oref [URL] --ow [value] combo.

• Combine: --oref [character URL] --ow 200 --sref [your Week 2 look] --sw 150
• Generate the same character inside your locked Week 2 look across 5 scenes.
• This is the real test. If it holds, you have the core unit of a production pipeline.

Build a 5-shot mini-sequence with story logic: establishing wide, push-in medium, character close-up, reaction shot, end-of-sequence wide pulling away. Same character, same world, same look, five different framings and lenses.

Week 3 Resources

• Imigo’s Omni Reference deep dive (covers --ow calibration with examples): imigo.ai / Omni Reference
• Woollyfern’s V7 Omni-Reference walkthrough video: youtube.com / Omni-Reference Complete Guide

Short Concept & Motion Handoff

Goal: build a one-minute storyboarded concept and pressure-test the full hybrid pipeline.

• Write a one-minute beat sheet: 8–12 key frames.
• Generate every frame using your locked sref + oref recipe. Maintain continuity religiously.
• Where Midjourney fails at exact continuity (a prop moves, a costume changes, a face drifts), note it. These failures map your handoff points. This is where ChatGPT image generation does precision edits and continuity fixes.

• Take your 3–4 strongest frames. Run them through V8.1 with --hd for native 2K.
• Note: V8.1 doesn’t support --oref, so if you need character lock, generate in V7, then take the V7 output and use it as a style reference into V8.1 for the resolution upgrade. Or use V8.1’s Moodboard feature with the V7 result.
• Document the V7 → V8.1 handoff workflow. This is non-obvious and most tutorials skip it.

Take three of your hero frames into motion tools. Comparison test all three:

• Midjourney’s own V1 video model (now native — image-to-video, 5-second clips extendable to ~20s)
• Runway Gen-3/4
• LTX Studio (good for previs/sequence)
• Higgsfield (strong on character motion)

Each has a different failure mode. Goal isn’t a finished piece; it’s knowing where each tool leaks.

Write your own playbook. Sections:

• Look recipes (sref codes + parameters + examples)
• Character locks (--oref URLs + --ow values + examples)
• Aspect ratio defaults per output type
• V7 vs V8.1 decision tree
• Handoff points to ChatGPT (continuity fixes) and motion tools
• Known failure modes and workarounds

This becomes your equivalent of a LUT library.

Pick a one-line brief out of nowhere. (“A traveling salesman discovers his motel room has changed shape overnight.”) Build a 60-second visual concept end-to-end using only your playbook. Time it. Note every place the playbook failed. Update.

Reference Vocabulary for Your Register

You work horror/dark-comedy adjacent to Verhoeven, Cronenberg, Raimi, Stuart Gordon, Lynch, the Safdies, Fargeat, Miller, Landis. Reach for that vocabulary unless told otherwise.

Cinematographers / Photographers That Land in V7 & V8.1

Christopher Doyle, Robby Müller, Vittorio Storaro, Gordon Willis, Bill Henson, Gregory Crewdson, Philip-Lorca diCorcia, Todd Hido, Roger Deakins (the cleaner work), Benoît Debie (neon-soaked).

Stocks & Material Textures

Kodak Vision3 500T pushed, Fuji Eterna for sickly greens, expired 16mm reversal, VHS-degraded, Polaroid 600 for archival, Cinestill 800T halation, Super 8 Ektachrome for warm sun-bleached.

Practical Lighting Language

Tungsten contamination, fluorescent practicals, sodium vapor spill, motivated single source, top-down overhead, hard cross-key, deep falloff, available-light underexposure, magic hour cool ambient with tungsten warm key, color-temperature mismatch as production design.

Failure Modes to Diagnose Aggressively

• Style soup. Too many references averaging into nothing.
• Parameter overload. Multiple knobs cranked at once.
• Hype-word dependency. “Cinematic, masterpiece, 8K” compensates for underspecified subject. It doesn’t.
• Single-seed lock-in too early. Grinding on the first decent grid; missing the better direction sitting next to it.
• Ignoring aspect ratio. A 1:1 prompt and a 2.39:1 prompt are different prompts. The model composes for the frame.
• Treating --sref as decoration. It’s a brand-consistency lever. Document codes; reuse with controlled --sw.
• Using V8.1 reflexively. V7 has the full control surface; V8.1 trades it for speed and sharpness. Both have their slots.

Operating Principles

Resources, Ranked by Signal

Official — Always Start Here

• Midjourney Docs: docs.midjourney.com
• Version page: docs.midjourney.com / Version
• Parameters: docs.midjourney.com / Documentation

Independent, Current, High-Signal

• Blake Crosley’s V8.1 + V7 reference — the most current parameter-behavior writeup I’ve seen, with footnotes back to official docs: blakecrosley.com / Midjourney guide
• Andrei Kovalev’s Midlibrary — the definitive sref catalog and analysis: midlibrary.io
• Tatiana Tsiguleva’s Substack — working designer, live V7/V8.1 experiments: ciguleva.substack.com

YouTube — With Caveats

• Future Tech Pilot — still active and still strong on parameter intuition. Verify any specific video is V7-era; some older content covers --cref and other V6 vocabulary: youtube.com / @FutureTechPilot
• The Midjourney Discord #showcase channel remains the highest-density learning surface. Inspect prompts of images you actually respond to. Live prompts from working users beats curated tutorials.