미드저니 프롬프트 엔지니어링

# Midjourney V7 Parameter Reference*Last Updated: February 2026*## Version Information- **V7 Release:** April 3, 2025- **Default Since:** June 17, 2025- **Video Model (V1):** June 2025---## What's Different in V7### Improved Capabilities- **Better prompt understanding** - Abstract concepts like "cinematic melancholy" or "ethereal nostalgia" work directly- **Smarter word weighting** - Beginning of prompt is weighted more heavily- **Simpler prompts work better** - V7 doesn't need the complexity V6 required- **Enhanced personalization** - On by default after 5-minute unlock process### Breaking Changes- **`--cref` is DEPRECATED** - Character reference no longer works- **Use `--oref` instead** - New "Omni Reference" system for characters/objects- **Style reference codes** - Old sref codes (pre-June 2025) need `--sv 4`- **Multi-Prompts (`::`) NOT COMPATIBLE with V7** - Separator syntax doesn't work in V7---## Prompt Structure### Key Insight: Simple Prompts Work BestMidjourney documentation emphasizes: **"Short, simple prompts work best."** V7 has much better natural language understanding than previous versions.### Recommended Order (Most to Least Important)```[Subject] > [Environment] > [Style/Mood] > [Lighting] > [Composition] > [Parameters]```Words at the beginning have more influence. Front-load your most important elements.### Seven Key Prompt Areas (from MJ docs)1. **Subject** - The main focus of your image2. **Medium** - Art style or technique (oil painting, photograph, 3D render)3. **Environment** - Where the scene takes place4. **Lighting** - How the scene is lit5. **Color** - Color palette and mood6. **Mood** - Emotional tone7. **Composition** - How elements are arranged---## Core Parameters### Aspect Ratio (`--ar`)| Use Case | Ratio | Parameter ||----------|-------|-----------|| Square (social, profile) | 1:1 | `--ar 1:1` || Standard photo | 3:2 | `--ar 3:2` || Widescreen | 16:9 | `--ar 16:9` || Ultrawide/cinematic | 21:9 | `--ar 21:9` || Portrait/mobile | 9:16 | `--ar 9:16` || Poster | 2:3 | `--ar 2:3` || Panoramic | 32:9 | `--ar 32:9` |### Stylize (`--s` or `--stylize`)Controls artistic interpretation vs prompt adherence. Range: 0-1000, Default: 100| Value | Effect ||-------|--------|| 0-50 | Maximum prompt adherence, minimal artistic liberty || 50-150 | Balanced (good for specific requirements) || 150-300 | Enhanced aesthetics (ideal for most creative work) || 300-600 | Strong artistic interpretation || 600-1000 | Maximum artistic freedom, prompt becomes suggestion |**Recommendation:** For commercial/professional work, stay between **100-250**.### Chaos (`--c` or `--chaos`)Controls variation between the 4 generated images. Range: 0-100, Default: 0| Value | Effect ||-------|--------|| 0 | All 4 images are similar || 20-40 | Moderate variety || 50-100 | High variety, exploratory |### Weird (`--weird`)Adds unusual/experimental aesthetics. Range: 0-3000, Default: 0**Note:** Weird is experimental and **not fully compatible with seeds** - results may vary even with same seed.| Value | Effect ||-------|--------|| 0 | Standard output || 50-150 | Subtle surrealism (good for artistic work) || 200-500 | Noticeably unusual || 500+ | Very experimental, unpredictable |### Quality (`--q` or `--quality`)**V7 values are different from previous versions!**| Value | GPU Time | Effect ||-------|----------|--------|| 1 | 1x | Default, standard quality || 2 | 2x | More detail and textures || 4 | 4x | Maximum detail (not compatible with --oref) |**Note:** Quality only affects the first set of 4 images. It doesn't affect variations, inpainting/outpainting, or upscales. For saving GPU time, use Draft Mode instead.### Style Raw (`--style raw` or `--raw`)In Standard Mode, Midjourney adds creative "auto-pilot" interpretation. Raw Mode turns this off.**With simple prompts:** More realistic, photo-like images**With detailed prompts:** Precise control over final lookUse for:- Cleaner, more controllable outputs- Photorealistic work- When you need precise control over aesthetics- Stylistically detailed prompts where you want exact interpretation### Draft Mode (`--draft`)- 10x faster generation- Half the GPU cost- Slightly lower quality- **Perfect for exploration/iteration before final renders****Workflow:**- Activate via the Draft Mode button in the Imagine bar, or add `--draft` to any prompt- When you like a draft, use **Enhance** button to regenerate at full quality (same composition, standard GPU time)- Draft images can also be directly upscaled without enhancing first- Works with Conversational Mode (AI writes prompts for you via text or voice)### Tile (`--tile`)Creates seamless, tileable patterns. Essential for:- Website backgrounds- Textures- Animated loops### Seed (`--seed`)Range: 0 to 4294967295. Same seed + same prompt = same result.**Use for:**- Iterating on a specific composition- Creating variations with small prompt changes- Testing how prompt changes affect output**Important Limitations:**- **Can't save styles** - Seeds only influence initial noise layout, not style/character. Use `--sref`, `--oref`, or `--p` for consistency- **Not always predictable** - Seeds may behave unexpectedly across different sessions. If you take a break and come back, results may differ- **Not always consistent** - Prompt, model version, and parameter changes can override seed effects- **Don't use with Turbo Mode** - Turbo focuses on speed; seed locking is unreliable### Repeat (`--repeat` or `--r`)Generate multiple image sets from a single prompt. Range: 2-40 (plan-dependent).| Plan | Max --repeat ||------|-------------|| Basic | 2 || Standard | 4 || Pro / Mega | 40 |**Use for:** Quickly generating many variations to pick from. Each repeat generates a full 4-image grid. Only works in Fast and Turbo modes.### Negative Prompt (`--no`)Excludes elements from generation:```--no text, words, letters, watermark, signature--no clutter, busy background--no people, faces```**⚠️ Moderation Warning:** MJ's moderation reads each word in `--no` independently. For example, `--no modern clothing` is interpreted as "no modern" AND "no clothing" — which can trigger content warnings. Instead of using `--no` to exclude clothing styles, include the specific clothing type you DO want in your prompt.---## Reference Systems### Quick Reference: When to Use Each Type| Type | Parameter | What It Does | Use When ||------|-----------|--------------|----------|| **Image Prompt** | (upload) | Influences content, composition, colors | You want similar subject/layout || **Style Reference** | `--sref` | Transfers visual style/vibe only | You want the aesthetic, not content || **Omni Reference** | `--oref` | Puts specific person/object into image | You need character/object consistency |### Image Prompts (Content Reference)Influences **content, composition, and colors** of your output. Upload an image and use it as inspiration.**Image Weight (`--iw`):** 0-3, default 1| Value | Effect ||-------|--------|| 0-0.5 | Subtle influence || 1 | Balanced (default) || 1.5-2 | Strong influence || 2-3 | Dominant influence |**Key insight:** Your text prompt should describe the **final image you want**, NOT instructions for changing the reference. MJ uses the image as inspiration, not as a source to modify.**Three ways to use:**1. Single image + text prompt2. Multiple images without text (blends them)3. Multiple images + text (combined guidance)### Style Reference (`--sref`)Transfers visual style from a reference image.**Basic:**```/imagine a mountain landscape --sref https://your-image-url.jpg```**With weight:**```/imagine a mountain landscape --sref https://url.jpg --sw 300```**Multiple references:**```/imagine a portrait --sref https://url1.jpg::2 https://url2.jpg::1 --sw 250```**Style Weight (`--sw`):** 0-1000, default 100. In V7, `--sw` has more impact with sref codes than with images.| Value | Effect ||-------|--------|| 0-50 | Subtle influence || 100 | Balanced || 200-400 | Strong style transfer || 500-1000 | Dominant style influence |**Incompatibilities:** `--sw` is not compatible with Moodboards.**Version:**- `--sv 6` (default) - Smarter style understanding- `--sv 4` - Required for sref codes created before June 2025- **Incompatibility:** `--sv` is not compatible with Moodboards.**Style Codes:** Use `--sref <numeric_code>` to apply a style from MJ's internal library instead of an image. You cannot create a style code from an uploaded image — codes come from the Style Explorer, Style Creator, or `--sref random`. See `rules/core-prompt-construction.md` for full style code documentation.**Random styles:**```/imagine a forest scene --sref random```**Behavior notes:**- Rerun/reroll and variations preserve the style code from the original prompt- `--sref random` with permutations or `--repeat` gives each image a different code### Omni Reference (`--oref`) -- REPLACES --crefUniversal reference for characters, objects, vehicles, creatures. **V7 only.****Basic:**```/imagine a person walking through Tokyo --oref https://character-image.jpg```**With weight (`--ow`):** 1-1000, default 100| Value | Effect ||-------|--------|| 25-50 | Subtle influence (style transfer) || 100 | Balanced preservation || 200-400 | Strong feature preservation || 400-1000 | Maximum fidelity to reference |**Best practice:** Keep `--ow` below 400 unless using very high stylize values.**Best Practices:**- Combine with clear text prompt - text is just as important- To change style: mention desired style at **START and END** of prompt- Example: "**Illustration** of a young woman with short gray hair **drawn by a comic book artist**"- Lower `--ow` when using style references- Intricate details (freckles, logos) may not match perfectly**Limitations (IMPORTANT):**- Costs 2x GPU time- NOT compatible with: Draft Mode, Fast Mode, inpainting, outpainting, `--q 4`- Results NOT compatible with: Vary Region, Pan, Zoom Out (use Editor instead)- Only ONE image can be used as Omni Reference per prompt---## Video/Animation### OverviewV1 Video Model (June 2025) enables image-to-video animation. This is NOT text-to-video -- you animate existing images.### Video Parameters| Parameter | Effect ||-----------|--------|| `--motion low` | Subtle ambient movement || `--motion high` | Dynamic movement || `--loop` | Seamless loop (end matches start) |### Duration- Base: 5 seconds- Extendable in ~4 second increments- Maximum: 21 seconds### Motion Prompting (Manual Mode)```Camera slowly pans left, clouds drift across sky, gentle ambient movement``````Subtle rotation, internal particles gently swirling, smooth morphing```### Technical Limitations- **Native resolution: 480p** (upscaling required for production)- **Frame rate:** ~24 fps- **Export formats:** MP4, GIF- **Cost:** ~8x GPU time of image generation### What Animates WellGood: Abstract patterns, atmospheric scenes (clouds, fog, water), flowing elements (hair, fabric, smoke), nature scenesBad: Detailed human faces (artifacts), complex mechanical movement, text or UI elements### Animation Workflow1. Generate base image with `--tile` for seamless edges2. Evaluate animation suitability3. Animate with low motion + loop for hero backgrounds4. Upscale to 1080p+ using Topaz, TensorPix, or DaVinci---## Effective Keywords### LightingVolumetric lighting, rim light, backlighting, golden hour, blue hour, dramatic shadows, soft diffused light, harsh directional light, bioluminescent, neon glow, ambient occlusion, subsurface scattering, caustics### Atmosphere/MoodEthereal, dreamy, nostalgic, melancholic, cinematic, epic, intimate, serene, mysterious, ominous, hopeful, whimsical### Quality/StyleHyperrealistic, photorealistic, stylized, octane render, redshift render, Cinema 4D, Houdini, professional photography, editorial, commercial, 8K render, ray traced### Camera/CompositionWide shot, extreme close-up, bird's eye view, shallow depth of field, bokeh, tilt-shift, rule of thirds, centered composition, macro, telephoto compression### Materials (for 3D aesthetics)Translucent, refractive, glossy, matte, subsurface scattering, volumetric, glass, chrome, ceramic, fabric---## Creating Text-Friendly CompositionsMidjourney doesn't understand "leave space for text" literally. Describe empty areas physically:**Effective:**```Subject positioned at far right, large empty area on left side``````Extreme wide shot with focal point in lower right corner``````Vast landscape with smooth gradient sky occupying upper two-thirds```**The invisible area technique:**```A photo with a huge invisible empty center area while the subject occupies the edges```---## Common Issues & Fixes| Problem | Solution ||---------|----------|| Too busy/cluttered | Add `--style raw`, lower `--s`, use `--no clutter, busy` || Not following prompt | Front-load key words, simplify prompt, lower `--s` || Unwanted text appearing | Add `--no text, words, letters, writing` || Too generic/stock-photo feel | Increase `--weird` slightly (50-150), add specific details || Subject in wrong position | Explicitly state position: "subject at far left edge" || Inconsistent style | Use `--sref` with consistent `--sw` value || Animation too jittery | Use `--motion low`, choose calmer source images || Old sref codes not working | Add `--sv 4` for pre-June 2025 codes || --cref not working | Switch to `--oref` (V7 change) |---## Text GenerationUse double quotation marks to render words in your images. Available in V6+.**Syntax:**```a neon sign that says "OPEN" in a dark alleya cartoon manual with the words "read the docs" in big text on the pages```**Best practices:**- Use **double quotes only** — single quotes and apostrophes don't work- Shorter words/phrases are more reliable than long sentences- Include context words like "with the words", "text", "written", "sign that says"- Works best with standard Latin alphabet (English letters)- If text is garbled: try `--style raw`, lower `--s`, or use the Editor/Vary Region to fix**Bad:** `a logo with the company name Boing` (no quotes)**Good:** `a pastel watercolor landscape with "imagine" written in the clouds`---## PermutationsGenerate multiple prompt variations from a single prompt using curly braces `{}`.**Syntax:**```a {red, green, yellow} bird        → generates 3 separate jobsa cat in a {forest, city, desert}  → generates 3 separate jobsa landscape --ar {1:1, 16:9, 2:3}  → tests 3 aspect ratios```**Works on any part of the prompt**, including parameters. Each permutation runs as a separate job consuming its own GPU time.**Plan limits:**| Plan | Max permutations per prompt ||------|----------------------------|| Basic | 4 || Standard | 10 || Pro / Mega | 40 |**Important:** Only works in Fast and Turbo modes. Not available in Relax mode.**Use cases for prompt engineering:**- A/B test keywords: `a portrait, {volumetric lighting, rim light, soft diffused light}`- Compare styles: `a cat, {oil painting, watercolor, pencil sketch} style`- Test parameters: `a landscape --s {50, 150, 300}`---## UpscalersV7 images start at 1024x1024 pixels (at 1:1). Upscaling doubles the dimensions.### Subtle vs Creative| Upscaler | What it does | Use when ||----------|-------------|----------|| **Subtle** | Doubles size, preserves original look as closely as possible | Final output is close to what you want || **Creative** | Doubles size, may add new details and make small changes | You want refinement or want to fix minor issues (hands, faces) |**Pro tip:** Creative upscale can sometimes correct small issues like awkward hands or odd expressions. You can re-run Creative upscale multiple times — each produces a slightly different result.### Resolution by Aspect Ratio| Aspect Ratio | Initial | After Upscale ||-------------|---------|---------------|| 1:1 | 1024 x 1024 | 2048 x 2048 || 4:3 | 1232 x 928 | 2464 x 1856 || 2:3 | 896 x 1344 | 1792 x 2688 || 16:9 | 1456 x 816 | 2912 x 1632 |**GPU cost:** Upscaling can take up to 2x the GPU minutes of the initial generation.**For larger sizes:** Use third-party upscalers (Topaz, TensorPix, DaVinci Resolve) after MJ upscale.---## Pan & Zoom Out### PanExpand the canvas in any direction to reveal new content beyond the original frame.**Use when:**- Composition is good but you need more space on one side (e.g., for text overlay)- You want to extend a scene without regenerating**Note:** Not compatible with images made using `--oref`. Use the Editor instead.### Zoom OutShrink the original image and fill in new surrounding content, creating a wider view of the scene.**Use when:**- Image is too tightly cropped- You want to reveal the environment around an existing subject- Building out a composition that started as a close-up**Note:** Not compatible with images made using `--oref`. Use the Editor instead.---## Artistic Mediums (from Official Docs)These medium keywords reliably produce distinct styles when used with V7. Tested format: `[medium] style [subject]`.### Print & TraditionalBlock Print, Ballpoint Pen Sketch, Cyanotype, Risograph, Ukiyo-e, Pencil Sketch, Watercolor, Pixel Art, Cross Stitch, Acrylic Pour, Cut Paper, Pressed Flowers, Oil Painting### SpecialtyBlacklight Painting, Paint-by-Numbers, Graffiti### Color ModesMillennial Pink, Acid Green, Sepia, Two Toned, Pastel, Duotone, Ebony, Neutral, CMYK, Indigo, Iridescent, Neon, Grayscale### Time Periods as StyleAdding a decade (e.g., "1920s", "1970s", "1990s") to your prompt shifts the entire visual style to match that era's aesthetic — lighting, color grading, composition, and subject presentation all adjust.---```--ar 16:9          Aspect ratio--s 200            Stylize (0-1000)--c 20             Chaos (0-100)--weird 100        Unusual aesthetics (0-3000)--style raw        Remove default beautification--draft            Fast/cheap iteration mode (10x faster, half cost)--q 2              Quality level (1=default, 2=2x GPU, 4=4x GPU)--no [items]       Exclude elements--seed 12345       Reproducible results (0-4294967295)--tile             Seamless/tileable pattern--p                Apply personalization (your style preferences)--sref [url]       Style reference--sw 300           Style weight (0-1000)--sv 4             Style version (for old codes)--oref [url]       Omni reference (characters/objects)--ow 200           Omni weight (0-1000)--iw 1.5           Image prompt weight (0-3)--repeat 4         Generate N image sets from one prompt (--r)--loop             Seamless video loop--motion low       Subtle animation"text here"        Render text in image (use double quotes){opt1, opt2}       Permutations (generates separate jobs)```---## Parameter Combinations### Maximum Control (Precise Requirements)```--s 50-100 --style raw --no [exclusions]```### Balanced Creative (Most Use Cases)```--s 150-250 --style raw```### Artistic Freedom (Exploratory)```--s 300-500 --weird 50-150```### Fast Iteration (Exploration Phase)```--draft --s 150```### Animation-Ready```--tile --s 150 --style raw```### Consistent Series (Using References)```--sref [url] --sw 200-300```---## Variations (Iteration Actions)### Subtle VariationsSmall, detailed changes to your image. **Use when:**- You like the overall image but want to tweak small areas- Fine-tuning without full redesign- Subject/composition is right but details need adjustment### Strong VariationsSignificant changes while keeping the main theme intact. **Use when:**- Exploring bold creative possibilities- Core idea is right but visual execution needs major push- Stuck and need fresh interpretation of same concept### Remix ModeAllows changing the prompt while making variations. Use when you want to vary AND adjust the prompt simultaneously.**Works with:** Variations, Vary Region, and Pan**Can change:** Both your prompt text AND parameters in a single operation---## Personalization (`--p`)Personalization applies the user's aesthetic preferences to generations. It learns from pair ranking votes and liked images on the Explore page.### Critical: ON by Default in V7**V7 is the first model with personalization enabled by default.** Users must unlock it (~5 minutes of ranking) before using V7, and it remains active unless explicitly toggled off. This means every V7 generation is influenced by the user's personal aesthetic profile.**Impact on pattern learning:** When personalization is on, MJ's output reflects a blend of prompt + user aesthetic. Patterns extracted from these sessions are partially user-specific, not purely universal MJ behavior. For reproducible prompt engineering, **disable personalization** to isolate prompt effects from profile effects.### Toggling On/Off- **On the website:** Click the personalization toggle button near the Imagine bar- **Per-prompt:** Add `--p` to force it on, or omit it (but the global toggle still applies)- **Profile code:** `--p xjspemh` applies a specific profile instead of the default### Profiles- Create multiple profiles with different styles (V7 Profiles and V6 Profiles are separate)- Each profile generates a unique ID/code- Stack multiple profiles: `--p 6odeoas 7enzken`- Set one as your default- Can also create **Moodboards** — curated image collections that define a specific style- **Note:** Personalization is "constantly in flux" — it changes subtly as you do more rankings, and the algorithm itself may update### Stylize + PersonalizationThe `--s` (stylize) parameter controls how much personalization is applied:- Lower stylize = Less personalization effect- Higher stylize = More personalization effect| --s Value | Personalization Effect ||-----------|----------------------|| 0 | Minimal personalization || 100 | Balanced (default) || 500+ | Strong personalization |---## Utility Tools### Describe (Vocabulary Discovery)Upload any image and MJ generates 4 creative prompt suggestions that could recreate it. **Essential for:**- Discovering effective vocabulary for visual styles- Learning how MJ interprets visual elements- Finding keywords you wouldn't have thought of- Translating visual analysis into MJ-effective language**Workflow tip:** Use Describe on reference images to discover vocabulary, then incorporate those keywords into your prompts.---## EditorThe Editor on midjourney.com provides tools for editing both Midjourney images and your own uploaded images.### Editor Tools (Light & Full Version)| Tool | Function ||------|----------|| **Undo/Redo/Reset** | Standard editing controls || **Move/Resize** | Move, scale, rotate image; change aspect ratio || **Paint (Erase)** | Select portions to regenerate || **Paint (Restore)** | Bring back erased areas || **Smart Select** | Create selection masks for areas to edit |### Edit Tab Only (Full Editor)| Tool | Function ||------|----------|| **Suggest Prompt** | Uses Describe to generate prompts from your image || **Custom Aspect Ratios** | Manually enter any aspect ratio || **Layers** | Add additional images to your composition || **Retexture** | Regenerate entire image in new style, keeping structure |### Export Options- **Upscale to Gallery** — Save to your MJ gallery- **Download Image** — Save to device (includes option for transparent PNG of erased areas)### Important Notes- **Omni Reference images:** Images made with `--oref` can only be opened in the Edit tab. You must remove the `--oref` and `--ow` parameters before submitting edits.- **Layers/Uploads:** Edited images using layers and uploaded images won't appear in Create/Organize pages unless you upscale them.- **Instead of Vary Region/Pan/Zoom Out** for `--oref` results, use the Editor for modifications.