Midjourney v7 Mastery Guide 2026: Parameters, Styles, and Pro Workflows

What's New in Midjourney v7 — Model Card and Capabilities

Midjourney v7 became the default model in early 2026 and represents the largest leap in the platform's history since v5. The headline change is a rebuilt diffusion backbone trained on a substantially larger and more curated dataset, paired with a new aesthetic predictor that scores compositions during generation rather than only at the end. The practical impact: fewer mangled hands, vastly improved typography rendering, photorealistic skin texture that no longer needs `--stylize 50` hacks, and prompt comprehension that finally handles multi-subject scenes without bleeding attributes between subjects. v7 also introduced a personalization layer (`--p`) that learns from your liked images across your account, which means two users typing the exact same prompt will now get different — but consistently in-style — results. Compared with v6, v7 produces approximately 35% sharper detail at default settings, supports native 2K rendering without the upscaler pipeline, and reduces the average iteration count to a usable image from roughly 5 generations to 2. The model card distributed by Midjourney lists default resolution at 2048x2048, native aspect ratio support from 1:5 to 5:1, and a token limit of 1500 characters for prompts (up from 1000 in v6). Style references and character references — which were experimental add-ons in v6 — are now first-class citizens with their own dedicated weight controls. The Draft Mode introduced in v6 was kept and accelerated; drafts now render in under two seconds and double as a cheap way to test compositions before committing to full quality renders. Importantly for working artists, v7 ships with a much stronger prior on photographic realism, meaning prompts that previously needed twenty keywords to escape Midjourney's signature 'painterly' look now respect 'photograph' as a top-level directive. The downside: v7's strong personalization can lock you into a visual rut. Many professional users deliberately disable `--p` for client work and re-enable it for personal exploration. If you're upgrading from v6, expect to retire roughly 60% of your old prompt library because the new model interprets style words like 'cinematic,' 'editorial,' and 'minimalist' with different defaults than v6 did.

Prompt Anatomy — The Six-Part Structure That Always Works

Every reliable Midjourney prompt can be decomposed into six layered components that the model reads roughly in order of weight: Subject, Medium, Style, Lighting, Composition, and Mood. Stack them in that order and you give the diffusion model an unambiguous brief. The Subject is the literal noun — 'a Bengal tiger,' 'a 1962 Jaguar E-Type,' 'a woman in her thirties with auburn hair.' Be specific; 'a tiger' produces an average tiger, 'a wet Bengal tiger emerging from monsoon water' produces a specific scene. The Medium answers 'what is this image?' — photograph, oil painting, charcoal sketch, 3D render, isometric vector, gouache illustration. Skipping the medium is the single biggest mistake new users make; v7 will default to a generic digital aesthetic if you don't specify. The Style names the artistic lineage — 'in the style of Saul Leiter,' 'shot like a Wes Anderson frame,' 'rendered like a Pixar production still,' 'inspired by Bauhaus poster design.' Be careful with living artists; Midjourney filters many names, and `--sref` is now the safer and more controllable route. The Lighting clause is what separates amateur prompts from professional ones — 'golden hour backlight,' 'soft north-facing window light,' 'dramatic Rembrandt lighting,' 'overcast diffused daylight,' 'neon-lit rain reflections,' 'studio beauty dish with white reflector.' Composition controls framing and lens — '85mm portrait, shallow depth of field,' 'wide-angle 24mm, deep focus,' 'overhead flat lay,' 'Dutch angle low shot,' 'centered symmetrical composition.' Mood ties the emotion together — 'melancholic,' 'triumphant,' 'serene,' 'tense,' 'nostalgic warmth.' A worked example: `Photograph of a Bengal tiger drinking from a monsoon puddle, low-angle DSLR shot, 200mm telephoto, golden hour rim light filtering through banana leaves, shallow depth of field with bokeh, cinematic muted tones, contemplative mood --ar 3:2 --style raw --v 7`. Notice the parameters come last, separated from the descriptive prose. This six-part structure isn't dogma — sometimes you'll skip mood, sometimes you'll merge style and medium — but it's a reliable scaffold you can lean on when a prompt isn't working. When a generation disappoints, audit which of the six layers is missing rather than throwing more adjectives at the problem.

Every Parameter, Explained

Midjourney parameters are flags appended at the end of the prompt with two dashes. Master these and you control 90% of what the model can do. `--ar` sets aspect ratio. Common values: `1:1` (square, default), `2:3` (portrait), `3:2` (landscape), `16:9` (widescreen), `9:16` (mobile/Reels/TikTok), `21:9` (cinema). v7 supports up to `5:1` panoramic. `--v` selects the model version. Defaults to v7 since February 2026, but you can still call `--v 6.1` or `--v 5.2` for legacy looks. `--style` modifies the aesthetic. The most important values: `--style raw` strips Midjourney's default 'enhancement,' giving you flatter, more photographic, more controllable results — use this for any photorealism work. `--style 4a`, `--style 4b`, `--style 4c` are legacy v4 sub-styles. For Niji, you have `--style original`, `--style cute`, `--style expressive`, `--style scenic`. `--chaos` (0-100, default 0) controls how much variation appears across the four-image grid. Low chaos means four similar compositions; high chaos means four wildly different interpretations. Use `--chaos 25` for exploration, `--chaos 0` when you've found your direction. `--weird` (0-3000, default 0) injects unusual aesthetic choices independent of chaos. `--weird 250` adds tasteful unexpectedness; `--weird 1500` produces avant-garde results; above 2000 things get genuinely strange. Chaos and weird interact, so don't crank both at once. `--stylize` or `--s` (0-1000, default 100) controls how much of Midjourney's artistic taste is applied. `--s 50` gives faithful literal renders; `--s 250` adds painterly flair; `--s 750+` produces highly stylized illustrations that may ignore parts of your prompt. For photorealism, stay between 50-150. `--no` is the negative prompt — `--no blur, text, watermark` tells the model to suppress those concepts. `--tile` generates seamlessly tileable patterns, useful for textures and backgrounds. `--seed` (0-4294967295) sets the random seed, letting you reproduce or iterate on a specific generation. To get a seed, react to a finished image with the envelope emoji in Discord, or check the metadata on the web app. `--quality` or `--q` (0.25, 0.5, 1, 2) trades GPU time for detail. Default 1 is balanced; `--q 2` is sharper but costs double; `--q 0.5` is fast and rough. `--niji 6` switches to the anime-specialized Niji model. `--p` enables personalization (requires you to have rated at least 200 images). `--repeat` (1-40) reruns the same prompt multiple times, useful for batch exploration. `--sref` and `--cref` deserve their own sections below.

Style References (--sref) — How to Build a Visual Library

Style references are the single most important workflow upgrade in v7. `--sref` lets you point Midjourney at one or more images and say 'render my prompt in that aesthetic.' The model extracts color palette, brushwork, lighting language, and compositional tendencies from the reference and applies them to your new subject without copying the content. Syntax: `--sref https://yourimage.com/ref.jpg`. You can stack multiple references: `--sref url1 url2 url3` averages their styles. You can weight them: `--sref url1::2 url2::1` gives the first reference double the influence. The `--sw` flag (style weight, 0-1000, default 100) controls how strongly the style transfers. `--sw 50` is a gentle nod toward the reference; `--sw 400` is heavy stylization; `--sw 800+` can override your prompt. Numerical style references (`--sref 12345`) point to Midjourney's internal style codes — there are around 4 billion possible codes, and the community has indexed thousands of the best ones at sites like SrefHunter and MJsref. A typical workflow: maintain a private Discord channel or Notion board of style references organized by use case — 'editorial fashion,' 'cozy interior,' 'cyberpunk,' 'children's book illustration,' 'corporate clean.' For client work, build a brand-specific sref library by uploading the client's existing marketing photography and noting which combinations produce on-brand outputs. When stacking references, three is usually the sweet spot; beyond five, the styles muddle into generic mid-tone slop. Style references work best when the reference image has a strong, distinct aesthetic. Murky reference images produce murky outputs. Style references do NOT transfer specific subjects — pointing at a photo of a red car will not make your output contain a red car, only inherit the photo's lighting and color grading. This is what separates `--sref` from `--cref` and `--iw` (covered next). One advanced trick: combine a numerical sref with an uploaded URL sref. `--sref 4209847 https://yourbrand.com/lookbook.jpg --sw 200` blends a known community style with your own brand reference.

Character References (--cref) — Consistent Characters Across Generations

If style references solved aesthetic consistency, character references solved the hardest problem in AI image generation: keeping the same person looking like the same person across dozens of images. `--cref` points Midjourney at a reference image of a person and instructs the model to preserve their face, hairstyle, and general appearance in new scenes. Syntax: `--cref https://yourimage.com/character.jpg`. The `--cw` flag (character weight, 0-100, default 100) controls fidelity. `--cw 100` preserves face, hair, AND clothing — useful when you want the character in the exact same outfit. `--cw 50` preserves face and hair but lets the model dress the character freely. `--cw 0` preserves only the face. For most narrative work — comic panels, brand mascots, recurring characters in a series — `--cw 50` is the right starting point. Character references work best when the source image is a clean, well-lit, front-facing portrait. Profile shots, partial faces, and group photos confuse the model. If you're generating an original character, the standard workflow is: generate a strong 'hero shot' of your character with a detailed prompt, upscale and download it, host it somewhere public (Discord works), then use that URL as your `--cref` for all subsequent generations. You can use multiple character references for two-character scenes: `--cref url1 url2` will attempt to keep both characters distinct. This works but is finicky — expect to roll several times before getting both faces right. Character references combine elegantly with style references. A common pro setup: `--cref https://yourchar.jpg --cw 50 --sref https://yourstyle.jpg --sw 200`. The character stays consistent while the entire scene's aesthetic shifts. Limitations: `--cref` is for human-like characters. For non-human characters (a specific dog breed, a fantasy creature design, a robot), `--cref` is unreliable; use `--sref` of your design language instead, or train a personalization layer. Also, `--cref` cannot perfectly preserve identity across extreme angle or expression changes — a character looking up dramatically may render with subtle facial drift. For mission-critical identity preservation (real client portraits in branded scenes), stack 3-4 reference images of the same person from different angles and rely on the strongest one as the cref while using the others as inspiration.

Image Weights (--iw) — Image Prompting and Remix

Beyond style and character references, you can use images themselves as part of your prompt. Drop an image URL at the start of your prompt and Midjourney treats it as compositional and content guidance, not just style. The `--iw` flag (image weight, 0-3 in v7, default 1) controls how much the image influences the output relative to your text. `--iw 0.25` means the text dominates and the image only nudges; `--iw 2` means the image is the primary driver and text is secondary; `--iw 3` is near-pure image-to-image transformation. Use cases: you want to take a sketch you drew and have Midjourney render it as a polished oil painting (`--iw 1.5` with appropriate medium prompts). You want to give the model a rough mood board and produce something thematically similar (`--iw 0.5`). You want to upgrade a low-quality reference photo into a high-resolution scene (`--iw 2`). Image prompts are different from `--sref` and `--cref` because they care about composition and subject matter, not just style or face. If your reference image shows a person standing in a doorway, an image prompt will produce something resembling a person in a doorway; an sref of the same image just transfers the color and mood. The Remix feature in the Midjourney web app is essentially a GUI for image prompting — you take an existing generation and adjust the prompt to vary it. Remix Strong (high `--iw` equivalent) keeps composition tight; Remix Subtle allows looser variation. A practical workflow for replacing stock photography: shoot a quick phone photo of the rough composition you want (a hand holding a coffee cup at a desk), use it as an image prompt with `--iw 1`, write a detailed prompt describing the upgraded scene ('professional product photography, soft natural window light, ceramic mug with steaming pour-over coffee, walnut wood desk, blurred plant in background, 50mm f/2.0, editorial mood'), and Midjourney will produce a finished image that matches your real composition but with elevated production value. This is faster and more controllable than describing the composition in words alone.

Multi-Prompts and Weight Syntax — The `::` Separator

Multi-prompts let you tell Midjourney that your prompt is actually a composition of distinct concepts, each with its own weight. The separator is a double colon `::`. Without weights, `space:: ship` is parsed as two equal-weight concepts: 'space' and 'ship.' Compare this to the single phrase 'space ship' which the model treats as one compound concept. Why does this matter? 'Bright wood' as a single phrase may produce wood that is bright in color; 'bright:: wood' tells the model these are two ideas — brightness and wood — and the result feels different. With explicit weights, syntax becomes powerful: `cyberpunk city::2 raining::1 neon signs::1 lonely figure::0.5`. The numbers are relative — what matters is the ratio. The first concept gets twice the weight of the next two and four times the last. Negative weights subtract concepts: `forest --no people` is one way to remove people, but `forest:: people::-1` is sometimes more surgical because it tells the model to actively render the inverse of 'people' rather than just suppress them. Negative multi-prompt weights work well for subtle removals: `portrait photograph:: smile::-0.5` produces a serious or neutral expression more reliably than `--no smile`. Multi-prompts also help with prompt bleed — the v6 problem where 'a man in a red shirt and a woman in a blue dress' would sometimes produce a man in blue and a woman in red. In v7, this is much less common, but for tricky multi-subject scenes you can write `man wearing red shirt::1 woman wearing blue dress::1` to force the model to treat each subject as a separate concept. The weight syntax is also useful for style blending: `oil painting::2 watercolor::1` produces an oil-dominant aesthetic with watercolor influences. Don't go wild — most prompts work fine without `::`. Use it when a specific compound phrase is being misread, when you need to force a weight imbalance, or when you want to use a negative concept weight rather than the `--no` parameter.

Negative Prompting (--no) — What to Suppress and Common Patterns

The `--no` parameter tells Midjourney to actively avoid concepts. Syntax: `--no blur, text, watermark, extra fingers`. Items are comma-separated. Common universal `--no` lists that working artists keep in a snippet: For photorealism: `--no cartoon, illustration, painting, 3D render, plastic skin, oversaturated, blurry, lowres, watermark, text, signature, frame, border`. For illustration: `--no photograph, photorealistic, 3D, render, blurry, lowres, watermark, text, signature, sketch lines, rough draft`. For clean product shots: `--no clutter, shadows, reflections, text, watermark, hands, fingers, people, extra props`. For portraits: `--no extra limbs, extra fingers, deformed hands, distorted face, blurry, lowres, plastic skin, watermark, text`. The `--no` parameter is most useful for systematic problems — Midjourney loves adding text and signatures to images that look like artwork, so `--no text, signature, watermark` is appended to almost every artistic prompt by experienced users. It's also useful for stripping unwanted defaults: if you ask for 'a forest' and keep getting paths through the forest, add `--no path, trail`. Negative prompting has limits. You cannot use `--no` to remove a concept that is central to your prompt — `tiger --no tiger` produces a confused image. You cannot use `--no` to enforce composition rules ('no centered subject' won't work — describe the composition you want instead). And in v7, the model's default outputs are clean enough that you usually need fewer `--no` items than you did in v6. Start with the universal list relevant to your medium, then add specific suppressions as patterns emerge in your iterations. One specific pattern: if you're getting unwanted text or gibberish words on signs, walls, or clothing in your generation, add `--no text, letters, words, writing, signage`. If you want text but it's coming out gibberish, switch to v7's improved text rendering by being explicit: `the words 'GRAND OPENING' painted in red on a wooden sign` — v7 handles short, specific text much better than v6 did.

Vary, Pan, Zoom — The Iteration Workflow

Once you have a generation you like, the upscale-and-iterate buttons turn a single image into a finished campaign. The button row below an upscaled image gives you: Vary (Strong), Vary (Subtle), Vary (Region), Zoom Out (1.5x, 2x, Custom), Pan (left/right/up/down), and the upscaler options (Subtle, Creative). Vary (Subtle) makes small adjustments while preserving composition — use it to refine a near-perfect image. Vary (Strong) re-rolls with bigger compositional changes — use it when you like the concept but want different framing. Vary (Region) is the killer feature. Click it, paint a mask over part of the image, then either accept the original prompt or modify it for just that region. Want to change just the character's outfit without touching the background? Mask the outfit, type 'wearing a leather jacket' in the new prompt box, generate. Want to remove a distracting object from a photo? Mask it, leave the prompt unchanged or describe what should be there instead. Vary (Region) is the inpainting tool that finally makes Midjourney usable for production retouching. Zoom Out extends the canvas beyond the original frame, having Midjourney imagine what's around your subject. Zoom 1.5x is gentle and usually believable; Zoom 2x is dramatic and can produce surreal results; Custom Zoom (up to 2x with a new prompt) lets you direct what fills the new area. Use Zoom to convert a tight portrait into a full environment shot, or to reframe a square image into a 16:9 landscape. Pan moves the canvas in one direction while keeping the original — use it to build out a panorama. Pan does not respect the original aspect ratio cleanly, so panoramas built this way may need cropping. The professional iteration loop: generate a 2x2 grid → upscale your favorite → Vary (Subtle) twice to refine → Vary (Region) on any remaining flaws → Zoom Out if you need more context → final upscale (Creative or Subtle depending on whether you want more detail invention or faithful enlargement). Most working artists report that 70% of their time is spent in this iteration loop, not in the initial prompt. Budget your effort accordingly.

Niji 6 — When to Use the Anime Model

Niji is Midjourney's specialized anime and illustration model, developed in collaboration with Spellbrush. Niji 6 (current as of 2026) is invoked with `--niji 6` and trained on a very different dataset than the main Midjourney model — it understands manga panel composition, anime lighting conventions, character design tropes, and Japanese illustration traditions in a way the main model simply does not. Use Niji 6 when you want: anime-style character art, manga panels, light novel cover illustrations, chibi designs, shounen action scenes, shoujo romantic compositions, mecha designs, JRPG concept art, vtuber character sheets, cute mascot designs, or any aesthetic descended from Japanese visual culture. Do NOT use Niji 6 for: photorealistic work (it cannot), Western comic book styles (use main MJ with style references), pure graphic design (use the main model), or any 'realistic' rendering. Niji 6 has its own style sub-flags: `--style original` for default anime, `--style cute` for chibi and kawaii aesthetics, `--style expressive` for more painterly, dynamic art, `--style scenic` for landscape-heavy compositions with smaller characters. Niji 6 supports all the main parameters — `--ar`, `--chaos`, `--stylize`, `--no`, `--sref`, `--cref` — and crucially, character references work very well with Niji because anime characters have more consistent feature 'language' than photographs. If you're making a recurring anime character for a series, generate a hero shot in Niji 6, then `--cref` that image for all subsequent scenes. Prompting Niji is different from prompting the main model. Niji responds well to anime-specific vocabulary: 'cel shaded,' 'screentone shading,' 'manga panel,' 'studio Ghibli style,' 'Kyoto Animation aesthetic,' 'sakuga frame,' 'shojo soft pastel palette.' It responds less well to photography vocabulary; saying '85mm lens' to Niji does little. Composition vocabulary still helps — 'dynamic low angle,' 'three-quarter view,' 'symmetrical character sheet' — but lighting language should be drawn from animation rather than photography: 'rim lighting against sunset sky,' 'soft anime lighting,' 'dramatic shadow contrast.' Niji is essential for anyone working in light novel covers, indie game art, vtuber assets, or animation pre-production. It is dead weight if you're shooting product photography or fashion editorial — use the main model.

Cross-Model Translation — Same Prompt, Different Model

Working AI artists in 2026 rarely commit to a single model. Midjourney has the strongest aesthetic taste but the worst prompt control. Flux (Pro and Dev) has near-perfect prompt adherence and superior typography but a less distinctive house style. Stable Diffusion XL (and its descendants like SD3 and Auraflow) gives total control via ControlNet, LoRAs, and local execution but demands technical setup. DALL-E 3 (now integrated into ChatGPT) handles natural language prompts conversationally and is best for users who don't want to learn parameter syntax. The same conceptual prompt looks different in each system. A working photo-style prompt — 'Photograph of a Bengal tiger drinking from a monsoon puddle, low-angle DSLR shot, 200mm telephoto, golden hour rim light filtering through banana leaves' — renders in Midjourney v7 with rich, painterly atmospheric depth and slightly enhanced color; the tiger feels heroic, almost mythological. The same prompt in Flux Pro renders with photographic literalism — the tiger looks like it was actually photographed, with neutral color science and accurate skin/fur texture; less mood, more truth. In SDXL with a photorealism LoRA, you get something between the two, with full control over seed, sampler, CFG, and steps via the local pipeline. In DALL-E 3, the prompt produces a competent illustration-leaning interpretation that is acceptable but visually unremarkable. The translation rules: when moving a prompt from Midjourney to Flux, drop most of the mood words and add more literal descriptors — Flux follows nouns better, Midjourney follows vibes better. When moving from Midjourney to SDXL, your `--ar` becomes a resolution choice, your `--stylize` becomes a CFG value (lower CFG = more creative, higher CFG = more literal), and your `--sref` becomes a LoRA or IP-Adapter reference. When moving to DALL-E, write the prompt as a paragraph in natural English rather than a parameter-stuffed comma list; DALL-E parses prose better than tag lists. The pro workflow many artists use: ideate fast in Midjourney v7 because its aesthetic is forgiving and produces shareable work quickly; when you need a specific composition, swap to Flux Pro for precise control; when you need to integrate the result with ControlNet/IP-Adapter pipelines for production work, finish in SDXL locally; use DALL-E inside ChatGPT for casual collaborative ideation when the client is in the room and you want a conversational interface. Treat the models as a stack, not a single choice.

Pro Workflows — Campaigns, Characters, Photoshoots, and Batch Variants

The four most common professional workflows in 2026, with concrete playbooks. Campaign mood boards: a brand wants 30 hero images across 5 product lines. Step one, build a style reference. Generate 20 exploratory images for the brand's general vibe using varied prompts. Pick the top 3, upscale them, host them, combine into `--sref url1 url2 url3 --sw 250`. Step two, lock the lighting language and color palette into a snippet you reuse across all prompts: e.g., 'soft north-window light, neutral warm tones, cream and walnut palette, editorial mood, 50mm f/2.8.' Step three, generate each product in the same template with only the subject swapped. Result: 30 visually cohesive images delivered in a single afternoon. Character consistency for a series: a creator is building a 12-episode illustrated podcast cover series with the same protagonist. Generate a hero portrait of the character with maximum detail and care, upscale, host. Use `--cref [url] --cw 75` for every cover. Use `--sref` to lock the illustration style separately. Each cover changes only the scene and pose; the character is preserved. For brand mascots, do the same but include the mascot's signature color in the prompt's text portion and use `--cw 100` to preserve outfit consistency. Photoshoot replacement: a startup needs lifestyle imagery for a landing page but cannot afford a real shoot. Workflow: write a casting brief ('woman in her early 30s with curly dark hair, warm smile, looks like a software engineer not a model'), generate 8 candidate hero portraits, pick one, upscale, save as `--cref`. Then list every scene needed ('working at a standing desk,' 'in a coffee shop with laptop,' 'on a phone call walking outside,' 'in a team meeting'). Generate each scene with the cref locked. Use Vary (Region) to fix any hand or face issues. Time to deliver: 2-3 hours for what would have been a $5,000 shoot. Caveat: disclose AI-generated people in any marketing where authenticity matters, and never use cref of a real person you don't have rights to. Batch variants for A/B testing: you have one concept that worked; you need 20 variations for ad testing. Use `--repeat 10` to fire 10 identical generations in one command. Use `--chaos 30` to introduce controlled variation. Combine with the Vary (Strong) button on your favorites. Export, tag, run them in your ad platform's split test, and let real CTR data tell you which aesthetic wins — your taste is biased, the data isn't. The common thread across all four workflows: lock your style reference and lighting language early, then iterate on subjects and compositions. The professionals don't write better prompts than amateurs — they reuse the same proven prompt scaffolding across hundreds of generations.

Frequently Asked Questions