Back to Details

midjourney-image

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Midjourney image production

Midjourney图像制作

Operating boundary

操作边界

Use this skill to prepare and supervise image work in Midjourney's documented consumer surfaces:
  • the Imagine bar, Create/Organize pages, Editor, Personalize page, and Moodboards page on 
    midjourney.com
    ;
  • the Midjourney Bot and documented slash commands in Discord.
Do not translate those buttons, slash commands, parameters, job IDs, or image URLs into a supposed REST endpoint, SDK, webhook, environment variable, or machine-to-machine contract. As verified 2026-07-09, Midjourney's Terms prohibit automated tools from accessing, interacting with, or generating through the Services, and its Community Guidelines say that, with rare explicitly granted exceptions, Midjourney does not provide an API or authorized third-party scripts. A Discord command is a user-interface action, not an API.
For an agent-safe workflow:
  1. Prepare the brief, prompts, parameter matrix, reference manifest, consent record, and QA checklist offline.
  2. Ask the authorized account holder to submit approved jobs manually in the official website or Discord UI.
  3. Have the user return downloaded outputs and, when useful, the job ID, seed, final prompt, version, and parameter record.
  4. Perform selection, comparison, rights review, compositing, typography, metadata, and QA in authorized local tools.
  5. Never drive the website or Discord with browser automation, self-bots, unofficial clients, scraping, macros, or simulated clicks. Never share an account.
If a customer says they have a rare written Midjourney automation exception, require the actual provider-issued agreement and its technical documentation. Treat that separately from this consumer-interface skill; do not infer authorization from a subscription, bot access, or an internal wrapper.
This skill is about still images. Midjourney documents an Animate creation action that can turn a suitable still into a five-second video, but that is a handoff to a video workflow, not an image-generation parameter recipe here.
本技能用于在Midjourney官方文档记载的面向消费者的界面中准备和监管图像制作工作：
  • midjourney.com
    上的Imagine栏、Create/Organize页面、Editor、Personalize页面和Moodboards页面；
  • Discord中的Midjourney机器人及官方记载的斜杠命令。
请勿将这些按钮、斜杠命令、参数、任务ID或图像URL臆想为REST端点、SDK、Webhook、环境变量或机器间交互协议。截至2026年7月9日核实，Midjourney服务条款禁止自动化工具访问、交互或通过服务生成内容，其社区准则也明确说明，除极少数明确授予的例外情况外，Midjourney不提供API或经授权的第三方脚本。Discord命令属于用户界面操作，而非API。
对于Agent安全的工作流：
  1. 离线准备需求简报、Prompts、参数矩阵、参考素材清单、授权记录和QA检查清单。
  2. 要求授权账户持有者在官方网站或Discord界面中手动提交已获批的任务。
  3. 让用户返回下载的输出文件，以及必要时的任务ID、seed、最终Prompt、版本和参数记录。
  4. 在授权的本地工具中完成筛选、对比、权利审核、合成、排版、元数据处理和QA工作。
  5. 绝对不要使用浏览器自动化、自机器人（self-bot）、非官方客户端、爬虫、宏脚本或模拟点击来操作网站或Discord。切勿共享账户。
如果客户声称自己拥有罕见的Midjourney自动化书面豁免权限，要求其提供官方出具的实际协议及技术文档。此类情况需与本消费端界面技能分开处理；不得通过订阅、机器人访问权限或内部封装工具推断授权有效性。
本技能仅针对静态图像。Midjourney官方记载了Animate创作功能，可将合适的静态图转换为5秒视频，但这属于视频工作流的范畴，并非本文档涵盖的图像生成参数方案。

Current capability map

当前能力图谱

The following snapshot is volatile and was verified against first-party documentation on 2026-07-09.
ChoiceUse it forProduction constraints
V8.1Current default general-image model; prompt fidelity, detail, SD/HD stillsReleased 2026-04-30 and default since 2026-06-10. 
--sd
is nominally 1024 px; 
--hd
nominally 2048 px. HD supports aspect ratios only through 4:1; SD through 14:1. The current compatibility chart does not support the Quality parameter in V8.1. The dedicated Draft page documents website-only V8.1 Draft, despite a conflicting mark in the Version chart. Turbo is not supported.
V7Omni Reference, established V7 personalization, and V7 Draft/Conversational productionOmni Reference forces V7 and costs about twice a normal image prompt. V7 quality 2/4 consumes roughly 2x/4x GPU time.
Niji 7Anime and Eastern illustration aesthetics, line work, literal/stylized graphic renderingSelect with the Niji model setting or 
--niji 7
. Image-weight range differs from V8.1/V7. Do not assume photoreal or brand-identity behavior.
V6.1Compatibility target for current Editor, Pan, Zoom, and region editing; legacy workflowsCurrent V8.1 edits may internally generate with V6.1, changing model character and detail. Character Reference belongs to V6/Niji 6; V7 uses Omni Reference.
V8.0 Alpha, V7, V6.1, and V6 remain lifecycle or compatibility choices, not aliases for the default. Record the actual model version on every approved asset. Old seeds, style codes, and prompt behavior are not a promise of future reproduction.
First-party documentation conflict: checked 2026-07-09, the Version compatibility chart marks Draft unavailable for V8.1, while the newer, feature-specific Draft and Conversational Modes page explicitly documents V8.1 website Draft with 24 previews at 512×512 for about 0.4 GPU minutes. Prefer the specific Draft page for an exploratory manual workflow, but verify the live website before scheduling or budgeting a batch and retain a non-Draft SD fallback.
以下内容为动态快照，已于2026年7月9日通过第一方文档核实。
选项适用场景生产限制
V8.1当前默认通用图像模型；Prompt保真度、细节表现、SD/HD静态图2026年4月30日发布，2026年6月10日起成为默认版本。
--sd
标称分辨率为1024像素；
--hd
标称分辨率为2048像素。HD仅支持最高4:1的宽高比；SD支持最高14:1。当前兼容性图表显示V8.1不支持Quality参数。尽管版本图表中有冲突标记，但专门的Draft页面记载了仅网站端可用的V8.1 Draft功能。不支持Turbo模式。
V7Omni Reference、成熟的V7个性化功能，以及V7 Draft/对话模式生产Omni Reference会强制使用V7，成本约为普通图像Prompt的两倍。V7的quality 2/4 分别消耗约2倍/4倍GPU时长。
Niji 7动漫与东方插画美学、线条绘制、写实/风格化图形渲染通过Niji模型设置或 
--niji 7
参数选择。图像权重范围与V8.1/V7不同。请勿默认其具备照片写实或品牌标识还原能力。
V6.1当前Editor、Pan、Zoom和区域编辑的兼容版本；遗留工作流当前V8.1的编辑操作可能在内部使用V6.1生成，会改变模型风格和细节。Character Reference属于V6/Niji 6的功能；V7使用Omni Reference。
V8.0 Alpha、V7、V6.1和V6仍为生命周期或兼容性选项，并非默认版本的别名。每个获批的资产都需记录实际使用的模型版本。旧的seed、风格代码和Prompt表现不保证未来可复现。
第一方文档冲突说明： 截至2026年7月9日核查，版本兼容性图表标记V8.1不支持Draft，但更新的、针对特定功能的Draft与对话模式页面明确记载了V8.1网站端Draft功能，可生成24张512×512预览图，消耗约0.4 GPU分钟。对于探索性手动工作流，优先参考专门的Draft页面说明，但在安排批量任务排期或预算前需核实网站实际情况，并保留非Draft的SD方案作为备选。

Resolution and edit lifecycle

分辨率与编辑生命周期

  • V8.1 SD jobs are documented at about 0.8 GPU minutes and HD at about 1.3 GPU minutes. V8.1 Draft produces 24 previews at 512×512 for about 0.4 GPU minutes.
  • HD is a generation mode, not merely a lossless export switch. An HD result cannot be upscaled further by the V8.1 upscalers.
  • Pan, Zoom, Editor, and Vary Region on a V8.1 image currently route through V6.1. Editing an HD image downscales the result to SD; upscale again only after the structural edit is approved.
  • The V8.1 Subtle and Creative upscalers take SD output to roughly 2× linear resolution. Creative may invent or alter detail; Subtle aims to preserve more closely. Neither guarantees pixel identity.
  • Resolution labels are nominal. Exact dimensions vary with aspect ratio, and the provider says HD dimensions may vary slightly.
  • 官方记载V8.1 SD任务耗时约0.8 GPU分钟，HD任务约1.3 GPU分钟。V8.1 Draft生成24张512×512预览图，耗时约0.4 GPU分钟。
  • HD是一种生成模式，而非单纯的无损导出开关。HD生成的结果无法通过V8.1的放大工具进一步放大。
  • 目前对V8.1图像执行Pan、Zoom、Editor和Vary Region操作时，会通过V6.1处理。编辑HD图像会将结果降级为SD；仅在结构编辑获批后再重新放大。
  • V8.1的Subtle和Creative放大工具可将SD输出提升至约2倍线性分辨率。Creative模式可能会新增或修改细节；Subtle模式则尽可能保留原有细节。两者均不保证像素完全一致。
  • 分辨率标签为标称值。实际尺寸会随宽高比变化，官方表示HD尺寸可能略有差异。

Access, capacity, and budget

访问权限、算力与预算

Plan facts below were verified 2026-07-09 and can change. Confirm the Manage Subscription page before committing money or a schedule.
PlanMonthly priceIncluded Fast timeRelax imagesImage concurrencyStealth
Basic$103.3 h / 200 minNo3 FastNo
Standard$3015 hUnlimited3 Fast or RelaxNo
Pro$6030 hUnlimited12 Fast or 3 RelaxYes
Mega$12060 hUnlimited12 Fast or 3 RelaxYes
Annual plans are advertised at a 20% discount paid for the year up front. Extra Fast time is listed at $4/hour. Unused subscription Fast time does not roll over.
  • Relax image jobs are unlimited on Standard, Pro, and Mega but may wait from 0–30 minutes depending on demand and prior use.
  • Turbo uses twice the Fast time and is experimental; it is unavailable in V8.1. A job can fall back to Fast when Turbo is unavailable.
  • Repeat and permutations are unavailable in Relax. Their per-submission caps are 4 jobs on Basic, 10 on Standard, and 40 on Pro/Mega; every expanded job consumes time.
  • Queued-job limits and running concurrency are not throughput guarantees. Schedule a small calibration batch before a deadline.
  • Use 
    /info
    in Discord or the website's account information to inspect remaining time and running/queued jobs; this is a manual UI check.
Budget by jobs and transformations, not only final images. Initial grids, Omni references, quality multipliers, upscales, variations, region edits, and remixes all consume GPU capacity. A defensible estimate states assumptions and keeps a reserve for repair.
以下套餐信息已于2026年7月9日核实，可能发生变动。在投入资金或确定排期前，请以“管理订阅”页面的信息为准。
套餐月费包含Fast时长Relax图像图像并发数Stealth
Basic$103.3小时 / 200分钟不支持3个Fast不支持
Standard$3015小时无限制3个Fast 或 Relax不支持
Pro$6030小时无限制12个Fast 或 3个Relax支持
Mega$12060小时无限制12个Fast 或 3个Relax支持
年付套餐宣传可享8折优惠，需一次性支付全年费用。额外Fast时长定价为每小时4美元。订阅包含的未使用Fast时长不会结转至下月。
  • Standard、Pro和Mega套餐的Relax图像任务无数量限制，但根据需求和过往使用情况，可能需要等待0-30分钟。
  • Turbo模式消耗两倍Fast时长，属于实验性功能；V8.1不支持。Turbo不可用时，任务会回退至Fast模式。
  • Relax模式不支持Repeat和排列组合（permutations）功能。每次提交的任务上限为：Basic套餐4个，Standard套餐10个，Pro/Mega套餐40个；每个扩展后的任务都会消耗时长。
  • 队列任务限制和运行并发数不代表吞吐量保证。在截止日期前先安排小批量校准测试。
  • 在Discord中使用 
    /info
    命令或通过网站的账户信息页面查看剩余时长和运行中/排队中的任务；这属于手动UI检查操作。
预算需按任务和转换次数计算，而非仅按最终图像数量计算。初始网格图、Omni参考图、质量倍率、放大、变体生成、区域编辑和Remix都会消耗GPU算力。合理的估算应明确假设条件，并预留修复余量。

Build a production brief before prompting

Prompt编写前先制作生产简报

Turn an aesthetic request into testable constraints:
  1. Deliverable: channel, crop, pixel need, number of finals, deadline, and whether exact text or logos are required.
  2. Subject invariants: identity, silhouette, product geometry, materials, wardrobe, pose, count, and prohibited changes.
  3. Scene: environment, time, weather, spatial relationship, and required negative space.
  4. Visual language: medium, era, palette, lighting, lens/composition language, texture, and finish.
  5. References: what each reference should control—content/composition, style, or identity/object consistency—and proof that the user may upload it.
  6. Risk: public visibility, confidentiality, real-person consent, trademarks, sensitive content, deception, and disclosure/provenance needs.
  7. Acceptance checks: framing, anatomy, text, product/logo fidelity, reference resemblance, seam quality, forbidden content, and technical export.
Separate hard constraints from preferences. Midjourney is generative, so a prompt cannot make exact geometry, identity, text, or logos deterministic. If an invariant must be exact, reserve it for photography, 3D, vector work, or post-compositing.
将美学需求转化为可测试的约束条件：
  1. 交付物： 发布渠道、裁剪要求、像素需求、最终稿数量、截止日期，以及是否需要精确的文字或Logo。
  2. 主体不变项： 身份特征、轮廓、产品几何形状、材质、服装、姿势、数量，以及禁止修改的内容。
  3. 场景： 环境、时间、天气、空间关系，以及所需的留白区域。
  4. 视觉语言： 媒介、年代、调色板、布光、镜头/构图语言、纹理和质感。
  5. 参考素材： 每个参考素材需控制的维度——内容/构图、风格，或是身份/物体一致性——以及用户有权上传该素材的证明。
  6. 风险： 公开可见性、保密性、真人授权、商标、敏感内容、误导性，以及披露/来源标注需求。
  7. 验收检查项： 构图、人体结构、文字、产品/Logo保真度、与参考素材的相似度、接缝质量、违禁内容，以及技术导出指标。
将硬性约束和偏好区分开。Midjourney是生成式工具，因此Prompt无法保证精确的几何形状、身份特征、文字或Logo完全确定。如果某项不变项必须完全准确，应留到摄影、3D、矢量制作或后期合成阶段实现。

Prompt construction

Prompt构建

Midjourney's documented prompt syntax is compact. Describe what should be visible, then place parameters at the end.
text
[subject and action], [environment], [medium/material], [composition and camera], [lighting], [palette/mood], [specific visible constraints] --parameters
This is a production heuristic, not mandatory prose. Prefer concrete visual nouns and modifiers over instructions about the model's thought process.
Midjourney官方记载的Prompt语法非常简洁。先描述应呈现的内容，再在末尾放置参数。
text
[subject and action], [environment], [medium/material], [composition and camera], [lighting], [palette/mood], [specific visible constraints] --parameters
这是生产实践中的经验法则，并非强制格式。优先使用具体的视觉名词和修饰词，而非针对模型思考过程的指令。

Prompt rules

Prompt规则

  • Put image prompt URLs at the beginning in Discord; on the website, add images through the image control and assign their role.
  • Put all parameters at the end, preceded by a space: 
    ... --ar 4:5 --s 150
    . Do not attach punctuation after a parameter value.
  • Keep the positive scene description coherent. Long lists of commands and negations can compete.
  • Specify count, placement, crop, and relationships explicitly: “one ceramic bottle centered, cap visible, label area blank.”
  • Treat camera terms as visual direction, not physically exact simulation. Pair a lens or shot phrase with distance, angle, depth, and composition.
  • For generated text, enclose a short Latin-alphabet phrase in double quotes. Single quotes are not the documented text syntax.
  • Save the submitted prompt, resulting prompt metadata, model, parameters, references, seed, and job ID before revision.
  • 在Discord中，将图像Prompt的URL放在开头；在网站端，通过图像控件上传图像并分配其作用。
  • 所有参数放在末尾，前面加空格：
    ... --ar 4:5 --s 150
    。参数值后不要加标点符号。
  • 保持正向场景描述的连贯性。过长的命令和否定词列表可能会相互冲突。
  • 明确指定数量、位置、裁剪和空间关系：例如“一个陶瓷瓶居中，瓶盖可见，标签区域空白”。
  • 将摄影术语视为视觉方向指引，而非物理精确模拟。将镜头或拍摄描述与距离、角度、景深和构图结合使用。
  • 如需生成文字，将短的拉丁字母短语用双引号括起来。单引号不是官方记载的文字生成语法。
  • 修改前，保存提交的Prompt、生成的Prompt元数据、模型、参数、参考素材、seed和任务ID。

Negative prompting

反向Prompt（Negative Prompting）

Use 
--no item1, item2
for visible exclusions. It is equivalent to negative multi-prompt weighting, not a semantic policy engine. Moderation evaluates terms independently, so 
--no modern clothing
may be interpreted as containing “no clothing.” Prefer a positive alternative such as “Victorian formalwear” and keep 
--no
to unambiguous objects like 
--no watermark, border
.
使用 
--no item1, item2
来排除不需要的可见元素。它等效于反向多Prompt加权，而非语义策略引擎。内容审核会独立评估每个词汇，因此 
--no modern clothing
可能被解读为包含“无衣物”的含义。优先使用正向描述替代，例如“维多利亚式正装”，仅将 
--no
用于明确的物体，如 
--no watermark, border
。

Reference strategy: give each image one job

参考素材策略：每张图只承担一个作用

Uploading an image never promises a copy. Crop references near the target aspect ratio and use clean, representative files. Supported upload formats documented for image prompts include PNG, GIF, WebP, JPG, and JPEG; the website upload limit is 10 MB.
上传图像绝不保证复制其内容。将参考素材裁剪至接近目标宽高比，使用清晰、有代表性的文件。官方记载的图像Prompt支持的上传格式包括PNG、GIF、WebP、JPG和JPEG；网站端上传大小限制为10MB。

Image Prompt — content, composition, and palette

图像Prompt（Image Prompt）—— 内容、构图和调色板

Use an Image Prompt when the reference should influence subject matter, composition, or colors.
  • Website: upload and choose Image Prompt.
  • Discord: place one or more direct image URLs before the text prompt.
  • Control influence with 
    --iw
    . The documented range is 0–3 for V8.1/V7 and 0–2 for Niji 7; default is 1.
  • One image generally needs a text prompt. Multiple images can be blended without text, though text usually makes the production intent auditable.
  • An image-only prompt cannot use 
    --stylize
    or 
    --weird
    .
Use a low-to-moderate 
--iw
first. If structure drifts, increase it; if the result merely imitates the reference and ignores the brief, reduce it or improve the text.
当参考素材需要影响主题、构图或色彩时，使用图像Prompt。
  • 网站端：上传后选择 Image Prompt。
  • Discord端：将一个或多个直接图像URL放在文本Prompt之前。
  • 通过 
    --iw
    控制影响程度。官方记载的范围为：V8.1/V7是0-3，Niji 7是0-2；默认值为1。
  • 单张图像通常需要搭配文本Prompt。多张图像可以在没有文本的情况下混合，但文本通常能让生产意图可追溯。
  • 纯图像Prompt无法使用 
    --stylize
    或 
    --weird
    参数。
先使用低到中等的 
--iw
值。如果结构偏离，就调高；如果结果只是模仿参考素材而忽略了需求简报，就调低或优化文本Prompt。

Style Reference — visual treatment, not depicted content

风格参考（Style Reference）—— 视觉处理风格，而非画面内容

Use Style Reference for color, medium, texture, lighting, and overall treatment while describing new content in text.
  • Website: assign the upload as a Style Reference. Discord: append 
    --sref URL
    .
  • Multiple style references are supported. Discord supports per-reference relative weights such as 
    --sref URL_A::2 URL_B::1
    .
  • --sw
    controls style influence from 0–1000; default 100.
  • A numeric style code can be used as 
    --sref code
    ; 
    --sref random
    explores random codes. A style image cannot be converted into a numeric code.
  • In V7, 
    --sv 6
    is the current default style-reference system and 
    --sv 4
    preserves the older pre-2025-06-16 behavior. Pin 
    --sv
    when migrating a style-sensitive campaign.
  • Moodboards are not compatible with 
    --sw
    or 
    --sv
    .
Do not expect the people, objects, or words in a style reference to transfer. Describe all required content in the text prompt.
当需要在文本描述新内容的同时，参考色彩、媒介、纹理、布光和整体处理风格时，使用风格参考。
  • 网站端：将上传的素材指定为 Style Reference。Discord端：追加 
    --sref URL
    。
  • 支持多个风格参考。Discord端支持为每个参考设置相对权重，例如 
    --sref URL_A::2 URL_B::1
    。
  • --sw
    控制风格影响强度，范围0-1000；默认值100。
  • 可以使用数字风格代码，格式为 
    --sref code
    ；
    --sref random
    可探索随机代码。风格图像无法转换为数字代码。
  • 在V7中，
    --sv 6
    是当前默认的风格参考系统，
    --sv 4
    保留了2025年6月16日之前的旧版行为。迁移对风格敏感的项目时，需固定 
    --sv
    值。
  • Moodboards与 
    --sw
    或 
    --sv
    不兼容。
不要期望风格参考中的人物、物体或文字会被迁移过来。在文本Prompt中描述所有需要的内容。

Omni Reference — one person, object, vehicle, or creature

Omni参考（Omni Reference）—— 单个人物、物体、车辆或生物

Use Omni Reference for stronger identity/object continuity across a new V7 scene.
  • It is V7-only; adding one Omni reference forces V7. Only one Omni image is accepted, and a text prompt is required.
  • Website: assign Omni Reference. Discord: append 
    --oref URL
    .
  • --ow
    ranges from 1–1000 and defaults to 100. Midjourney recommends staying below 400 in typical work unless intentionally combining with strong stylization.
  • It costs roughly twice a standard image prompt.
  • It is compatible with Personalization, moodboards, 
    --stylize
    , Style References, and Image Prompts.
  • It is not compatible with Draft, Conversational mode, Fast mode, 
    --q 4
    , or direct V6.1-based Pan/Zoom/Vary Region editing.
To edit an Omni result, open it in the Editor and remove the Omni image and 
--oref
/
--ow
; expect continuity to weaken because the edit currently uses V6.1. Exact faces, hands, small emblems, logos, and product details remain a QA problem.
Only upload a real person's image with the necessary consent and rights. Do not create sexualized, abusive, derogatory, inflammatory, deceptive, or otherwise disallowed manipulations. A blocked job is not evidence that the underlying request would have been permitted.
当需要在新的V7场景中保持更强的身份/物体连续性时，使用Omni参考。
  • 仅支持V7；添加一个Omni参考会强制使用V7。仅接受一张Omni图像，且必须搭配文本Prompt。
  • 网站端：指定为 Omni Reference。Discord端：追加 
    --oref URL
    。
  • --ow
    范围为1-1000，默认值100。Midjourney建议在常规工作中保持在400以下，除非有意结合强烈的风格化效果。
  • 成本约为标准图像Prompt的两倍。
  • 兼容个性化（Personalization）、Moodboards、
    --stylize
    、风格参考和图像Prompt。
  • 不兼容Draft、对话模式、Fast模式、
    --q 4
    ，以及基于V6.1的直接Pan/Zoom/Vary Region编辑。
如需编辑Omni生成的结果，在Editor中打开并移除Omni图像和 
--oref
/
--ow
参数；由于当前编辑使用V6.1，连续性会有所减弱。精确的面部、手部、小型徽章、Logo和产品细节仍属于QA问题范畴。
仅在获得必要授权和权利的情况下，才可上传真人图像。不得创建性化、辱骂性、贬损性、煽动性、欺骗性或其他被禁止的内容。被拦截的任务不代表底层请求本来会被允许。

Character Reference — legacy compatibility

角色参考（Character Reference）—— 遗留兼容性

--cref
and 
--cw
belong to V6/Niji 6 workflows. 
--cw
ranges 0–100 and defaults to 100. For a V7 workflow, migrate to Omni Reference; do not silently keep 
--cref
. Neither system guarantees an exact real person or exact logos/details.
--cref
和 
--cw
属于V6/Niji 6的工作流。
--cw
范围为0-100，默认值100。对于V7工作流，请迁移至Omni Reference；不要静默保留 
--cref
。两种系统都不保证完全还原真人或精确的Logo/细节。

Personalization and Moodboards

个性化（Personalization）与Moodboards

Personalization applies a learned profile with 
--p
; the global V7/V8 profile must first be unlocked by selecting images. Profiles and moodboards have IDs that resolve to versioned codes. Save the resolved code for campaign reproducibility rather than relying only on a mutable profile ID.
Moodboards are broader curated aesthetic guides and work with V6, V7, and V8.1. They can be selected in the web UI or used as 
--p mID
/the resolved code. 
--stylize
controls their strength, but moodboards cannot be combined with 
--sw
or 
--sv
. Do not stack Personalization, a moodboard, and style references reflexively; isolate one influence at a time during calibration.
个性化功能通过 
--p
参数应用训练好的配置文件；全局V7/V8配置文件需先通过选择图像解锁。配置文件和Moodboards都有对应的ID，可解析为带版本的代码。保存解析后的代码以保证项目可复现，不要仅依赖可变的配置文件ID。
Moodboards是更广泛的策划美学指南，支持V6、V7和V8.1。可以在网页UI中选择，或通过 
--p mID
/解析后的代码使用。
--stylize
控制其强度，但Moodboards不能与 
--sw
或 
--sv
结合使用。不要盲目叠加个性化、Moodboard和风格参考；校准阶段每次只测试一种影响因素。

Parameter decisions

参数决策

These are documented controls, not universal “best settings.” Change one material variable at a time during diagnosis.
ParameterCurrent documented behaviorProduction use and caution
--ar W:H
Aspect ratio; default 1:1; integer ratiosMatch the delivery crop at generation time. V8.1 limits: SD through 14:1, HD through 4:1. Extreme ratios are experimental and ratio is not pixel size.
--chaos N
/ 
--c N
0–100, default 0Increase breadth among a grid; higher values reduce predictability and can weaken adherence. Use higher chaos for ideation, lower for controlled revisions.
--stylize N
/ 
--s N
0–1000, default 100Higher values increase Midjourney's aesthetic influence and may reduce literal adherence. Lower it for layout, product, or text-sensitive work.
--weird N
/ 
--w N
Experimental, 0–3000Explore unusual aesthetics. Behavior can change and is not fully compatible with deterministic seed comparisons.
--quality N
/ 
--q N
V7 accepts 1, 2, 4; q2/q4 costs about 2x/4x. The current V8.1 chart marks Quality unsupported.In V7, spend quality after composition is promising. Quality affects the initial job, not later variations, inpainting, outpainting, or upscaling. V7 
--q 4
is incompatible with Omni. Do not copy a V7 quality flag into V8.1.
--seed N
Integer 0–4,294,967,295Useful for controlled comparisons, not archival determinism. V8.1 is documented as highly repeatable under like conditions, but seeds can drift across sessions, models, settings, and Turbo.
--tile
Makes a single seamless repeating tilePreview the repeated pattern externally. Upscaling can break the seam; test the exact exported asset.
--no ...
Negative multi-prompt influenceKeep exclusions unambiguous; moderation parses each word. Positive replacement is often safer.
--raw
Reduces default Midjourney styling/automationUseful for literal, photographic, layout, or text-sensitive work; combine with precise direction, not vague prompts.
--iw N
Image Prompt influenceV8.1/V7 0–3; Niji 7 0–2; default 1. Tune against text adherence.
--sref
, 
--sw
, 
--sv
Style image/code, strength, system versionPin references and 
--sv
for style migration. 
--sw
default 100, range 0–1000.
--oref
, 
--ow
V7 Omni image and strengthOne image, text required, default 100, range 1–1000, about 2x GPU, compatibility limits.
--p
Personalization profile or moodboardSave resolved codes because profiles evolve.
--draft
Draft generationV8.1 website Draft is a 24-image 512×512 exploration board; V7 Draft uses half standard GPU cost and generates four images. Enhance or vary a selected concept before judging finish.
--fast
, 
--relax
, 
--turbo
Per-job speed modeRelax availability is plan-limited; repeat/permutations do not work in Relax; Turbo costs 2x and does not support V8.1.
--repeat N
/ 
--r N
Repeat a promptFast/Turbo only; plan caps 4/10/40. Treat every repeat as a paid job.
{a,b,c}
Permutation syntaxExpands prompt/parameter variants in Fast/Turbo only. Confirm the job count before submission. Escape a literal comma inside braces with 
\,
.
--stealth
, 
--public
Per-job visibility on the websitePro/Mega only for Stealth. Does not hide a creation from people in a shared Discord channel.
--v
, 
--niji
, 
--sd
, 
--hd
Model family and V8.1 resolutionPin them in a production record. Omni can force V7. HD limits editing/upscaling and aspect ratio.
The live Parameter List also includes video-only controls. Do not copy 
--motion
, 
--loop
, 
--end
, 
--bs
, or 
--video
into a still-image prompt merely because they appear in the same documentation.
以下是官方记载的控制项，并非通用的“最佳设置”。排查问题时，每次只改变一个核心变量。
参数当前官方记载的行为生产使用注意事项
--ar W:H
宽高比；默认1:1；整数比例生成时匹配交付裁剪要求。V8.1限制：SD最高支持14:1，HD最高支持4:1。极端比例属于实验性功能，比例不等于像素尺寸。
--chaos N
/ 
--c N
0-100，默认0增加网格图的多样性；值越高，可预测性越低，且可能减弱对需求的遵循度。创意发散阶段使用较高的chaos值，受控迭代阶段使用较低值。
--stylize N
/ 
--s N
0-1000，默认100值越高，Midjourney的美学影响越强，可能降低对字面描述的遵循度。对于布局、产品或文字敏感的工作，调低该值。
--weird N
/ 
--w N
实验性功能，0-3000探索非常规美学。其行为可能变化，且与确定性的seed对比不完全兼容。
--quality N
/ 
--q N
V7支持1、2、4；q2/q4成本约为2倍/4倍。当前V8.1兼容性图表标记不支持Quality参数。在V7中，当构图符合预期后再提高quality值。Quality仅影响初始任务，不影响后续的变体生成、局部重绘、外扩或放大。V7的 
--q 4
与Omni不兼容。不要将V7的quality参数直接用于V8.1。
--seed N
整数，范围0–4,294,967,295适用于受控对比，而非归档级的确定性复现。官方记载V8.1在相同条件下具有高度可重复性，但seed在不同会话、模型、设置和Turbo模式下可能发生偏移。
--tile
生成单个无缝重复图案在外部工具中预览重复效果。放大可能会破坏接缝；需测试实际导出的资产。
--no ...
反向多Prompt影响排除项需明确无歧义；内容审核会逐词解析。正向替代通常更安全。
--raw
减少Midjourney默认的风格化/自动化效果适用于写实、摄影、布局或文字敏感的工作；需搭配精确的指引，而非模糊的Prompt。
--iw N
图像Prompt影响强度V8.1/V7范围0-3；Niji 7范围0-2；默认值1。根据文本遵循度调整。
--sref
, 
--sw
, 
--sv
风格图像/代码、强度、系统版本迁移风格时固定参考素材和 
--sv
值。
--sw
默认值100，范围0-1000。
--oref
, 
--ow
V7 Omni图像及强度仅支持单张图像，需搭配文本，默认值100，范围1-1000，约2倍GPU消耗，存在兼容性限制。
--p
个性化配置文件或Moodboard保存解析后的代码，因为配置文件会更新。
--draft
Draft生成V8.1网站端Draft是24张512×512图像的探索板；V7 Draft消耗标准GPU成本的一半，生成4张图像。在判断完成度前，先对选中的概念进行增强或变体生成。
--fast
, 
--relax
, 
--turbo
单任务速度模式Relax的可用性受套餐限制；Repeat/排列组合在Relax模式下不可用；Turbo成本为2倍，不支持V8.1。
--repeat N
/ 
--r N
重复生成Prompt仅支持Fast/Turbo模式；套餐上限为4/10/40。每次重复都视为付费任务。
{a,b,c}
排列组合语法仅在Fast/Turbo模式下扩展Prompt/参数变体。提交前确认任务数量。大括号内的字面逗号需用 
\,
转义。
--stealth
, 
--public
网站端单任务可见性Stealth仅支持Pro/Mega套餐。不会向共享Discord频道的成员隐藏创作内容。
--v
, 
--niji
, 
--sd
, 
--hd
模型系列和V8.1分辨率在生产记录中固定这些参数。Omni可能强制使用V7。HD会限制编辑/放大和宽高比。
实时更新的参数列表中还包含仅适用于视频的控制项。不要仅仅因为它们出现在同一份文档中，就将 
--motion
、
--loop
、
--end
、
--bs
或 
--video
用于静态图像Prompt。

Draft and Conversational modes

Draft与对话模式

Use Draft to search a concept space cheaply, not to approve fine detail.
使用Draft可以低成本探索概念空间，但不要用于审批精细细节。

V8.1 Draft

V8.1 Draft

On the website, Draft creates 24 square 512×512 previews. Select candidates, then Vary or Remix them to SD/HD. It is ideal for composition, palette, silhouette, and art-direction search, but too coarse for typography, fingers, packaging, textures, or delivery QA.
在网站端，Draft会生成24张512×512的方形预览图。选择候选方案后，通过Vary或Remix生成SD/HD版本。它非常适合用于构图、调色板、轮廓和艺术方向探索，但对于排版、手指、包装、纹理或交付QA来说过于粗糙。

V7 Draft

V7 Draft

V7 Draft returns four images at about half standard GPU cost. Use Enhance to generate a higher-quality interpretation. Enhancement can change details, so re-check all invariants.
V7 Draft生成4张图像，消耗约标准GPU成本的一半。使用 Enhance 生成更高质量的版本。增强操作可能会改变细节，因此需重新检查所有不变项。

Conversational mode

对话模式（Conversational mode）

Conversational mode is a website prompt-writing assistant available for V7/V8.1. It accepts typed or spoken natural language and can rewrite a request into a prompt. Typed conversation can work with or without Draft; voice requires Draft. Treat its rewritten prompt as an editable proposal: inspect the actual submitted text, parameters, version, and references. It is not an autonomous API or a substitute for rights/safety review.
对话模式是网站端的Prompt编写助手，支持V7/V8.1。它接受输入或语音的自然语言，可将请求重写为Prompt。文本对话可搭配或不搭配Draft使用；语音对话需要Draft。将其重写的Prompt视为可编辑的提案：需检查实际提交的文本、参数、版本和参考素材。它不是自主API，也不能替代权利/安全审核。

Controlled iteration and repair

受控迭代与修复

Use a funnel rather than changing the entire prompt after every result.
  1. Explore: choose the model and target aspect ratio; run a Draft or low-cost standard calibration with moderate settings.
  2. Select structure: judge silhouette, layout, pose, crop, and negative space before surface detail.
  3. Lock variables: save prompt, model, parameter set, references, resolved profile/style codes, job ID, and seed.
  4. Vary: use Subtle when the composition is right and Strong when the concept needs movement. Variations are related generations, not exact edits.
  5. Remix: change prompt text or compatible parameters while making a variation/edit. On web, Remix populates the Imagine bar or Editor; in Discord, enable Remix in settings or 
    /prefer remix
    . Changing aspect ratio via Remix stretches rather than outpaints—use Pan, Custom Zoom, or Editor to add canvas.
  6. Repair locally: use Editor erase/restore, Smart Select, layers, Pan, Zoom, or region editing. Mask enough context for the model to solve the boundary. Current edits route through V6.1, so compare identity and style again.
  7. Retexture: use when the structure is acceptable but the whole treatment must change. It regenerates the full image while retaining broad structure; it is not a pixel-preserving material swap.
  8. Upscale last: choose Subtle for preservation or Creative for invented detail. Run a difference/side-by-side check.
  9. Finish externally: composite exact typography, logos, legal copy, product geometry, or color-managed assets in a suitable editor.
采用漏斗式迭代方式，不要每次得到结果后就修改整个Prompt。
  1. 探索： 选择模型和目标宽高比；使用中等设置运行Draft或低成本的标准校准任务。
  2. 筛选结构： 先判断轮廓、布局、姿势、裁剪和留白，再看表面细节。
  3. 锁定变量： 保存Prompt、模型、参数集、参考素材、解析后的配置/风格代码、任务ID和seed。
  4. 变体生成： 构图合适时使用Subtle模式，概念需要调整时使用Strong模式。变体是相关的生成结果，而非精确编辑。
  5. Remix： 在生成变体/编辑时，修改Prompt文本或兼容的参数。在网页端，Remix会自动填充到Imagine栏或Editor中；在Discord中，需在设置中开启Remix或使用 
    /prefer remix
    命令。通过Remix修改宽高比会拉伸图像而非外扩画布——如需增加画布，请使用Pan、Custom Zoom或Editor。
  6. 局部修复： 使用Editor的擦除/恢复、Smart Select、图层、Pan、Zoom或区域编辑功能。提供足够的上下文蒙版，让模型能够处理边界。当前编辑通过V6.1处理，因此需重新对比身份特征和风格。
  7. 重纹理（Retexture）： 当结构可接受但整体风格需要完全改变时使用。它会重新生成整张图像，同时保留大致结构；并非保留像素的材质替换。
  8. 最后放大： 选择Subtle模式保留细节，或Creative模式新增细节。进行差异对比/并排检查。
  9. 外部收尾： 在合适的编辑器中合成精确的排版、Logo、法律文本、产品几何形状或色彩管理后的资产。

Editor privacy nuance

Editor隐私细节

The current Editor documentation makes a specific visibility distinction. Ordinary Editor creations are saved in the Editor and do not appear in Create/Organize unless upscaled. When a workflow uses externally uploaded images in the Editor or multiple layers, however, both the uploads and edited results—including upscales—are documented as visible only to that user, absent from other users on midjourney.com, and unavailable to Discord's 
show
command. This is a useful Basic-plan non-public route for a genuine external-image editing job; it is not a general privacy switch for Imagine, Image Prompt, Style Reference, or Omni jobs.
Do not equate this product-visibility behavior with contractual confidentiality or local processing. Midjourney still receives, stores, and processes the content; its Terms, Privacy Policy, broad content license, external-image rules, and account security still apply. Require client/data-policy approval before uploading confidential material, minimize the input, and prefer a local or contractually approved system when non-disclosure is essential.
当前Editor文档对可见性有明确的区分。普通的Editor创作会保存在Editor中，除非放大，否则不会出现在Create/Organize页面。但当工作流在Editor中使用外部上传的图像或多个图层时，上传的素材和编辑后的结果（包括放大后的版本）官方记载为仅对该用户可见，其他midjourney.com用户无法查看，也无法通过Discord的 
show
命令访问。对于真实的外部图像编辑任务，这是Basic套餐可用的非公开途径；但它不是Imagine、图像Prompt、风格参考或Omni任务的通用隐私开关。
不要将这种产品可见性行为等同于合同保密性或本地处理。Midjourney仍会接收、存储和处理内容；其服务条款、隐私政策、广泛的内容许可、外部图像规则和账户安全条款仍然适用。上传机密材料前需获得客户/数据政策批准，尽量减少输入内容，当保密要求较高时，优先使用本地或经合同认可的系统。

Typography and graphic design

排版与平面设计

Midjourney supports short generated text in V6 and later, with these documented tactics:
  • put a short Latin-alphabet phrase in double quotes;
  • use 
    --raw
    and/or reduce 
    --stylize
    when letter accuracy matters;
  • repair with Editor/Vary Region when the concept is good;
  • simplify the scene so text has a clean surface and clear hierarchy.
Generated typography is still probabilistic. For a brand name, trademark, price, URL, accessibility text, legal copy, medical label, packaging facts, or any phrase where one character matters, generate a blank or placeholder area and composite approved text afterward. QA spelling, punctuation, glyph shape, reading order, contrast, safe area, and export resolution. Do not describe a Midjourney-rendered logo as an exact brand asset.
Midjourney在V6及更高版本中支持生成短文本，以下是官方记载的技巧：
  • 将短的拉丁字母短语放在双引号中；
  • 当字母准确性很重要时，使用 
    --raw
    和/或降低 
    --stylize
    值；
  • 当概念符合预期时，用Editor/Vary Region修复；
  • 简化场景，让文字有干净的背景和清晰的层级。
生成的排版仍然是概率性的。对于品牌名称、商标、价格、URL、无障碍文本、法律文本、医疗标签、包装说明或任何单个字符都很重要的短语，先生成空白或占位区域，之后再合成经批准的文本。QA需检查拼写、标点、字形、阅读顺序、对比度、安全区域和导出分辨率。不要将Midjourney生成的Logo描述为精确的品牌资产。

Privacy, rights, safety, and provenance

隐私、权利、安全与来源

Public-by-default and Stealth

默认公开与Stealth

Midjourney is open by default. Creations can appear on Explore even when generated in a private Discord server or direct message unless Stealth is active.
  • Stealth is available only on Pro and Mega. It controls visibility on the Midjourney website.
  • Stealth does not hide messages or images from participants in a shared Discord channel. For the strongest available privacy, use Stealth with the web Create page, a direct message, or a private server.
  • Turning Stealth on does not retroactively privatize existing creations; make them private individually or in bulk.
  • Trashing a creation is not the same as making it private; trashed work may still be discoverable.
  • Private content stays private after downgrade, but the Terms describe Stealth as a best-efforts non-publication commitment, not an absolute confidentiality guarantee.
  • Making a private image a profile image, banner, or spotlight can make it—and related grid images—public.
Before uploading unreleased products, client work, personal data, or trade secrets, compare this model with the organization's data policy. The Privacy Policy says Midjourney collects information including text/image prompts, public chats, account/contact data, IP address, and billing-related information. Discord separately controls Discord data. Account/data deletion is initiated through the account page, cancels the subscription, has a seven-day reversal window, and is documented as completing within 30 days, subject to policy/legal exceptions.
Midjourney默认公开。即使是在私人Discord服务器或私信中生成的创作，也可能出现在Explore页面，除非开启了Stealth模式。
  • Stealth仅支持Pro和Mega套餐。它控制Midjourney网站上的可见性。
  • Stealth不会向共享Discord频道的参与者隐藏消息或图像。要获得最强的隐私保护，将Stealth与网页Create页面、私信或私人服务器结合使用。
  • 开启Stealth不会将现有创作追溯设为私有；需单独或批量将其设为私有。
  • 删除创作不等于将其设为私有；被删除的作品仍可能被找到。
  • 降级后私有内容仍保持私有，但服务条款将Stealth描述为尽力而为的不公开承诺，而非绝对的保密保证。
  • 将私有图像设为头像、横幅或精选内容，可能会使其——以及相关的网格图——变为公开。
上传未发布产品、客户作品、个人数据或商业机密前，请将此模式与组织的数据政策进行对比。隐私政策规定，Midjourney收集的信息包括文本/图像Prompt、公开聊天、账户/联系数据、IP地址和账单相关信息。Discord数据由Discord单独控制。账户/数据删除需通过账户页面发起，会取消订阅，有7天的撤销期，官方记载完成时间为30天内，受政策/法律例外情况约束。

Rights and consent

权利与授权

Under the Terms effective 2026-05-27, a compliant user owns Assets to the fullest extent possible under applicable law, subject to third-party rights and stated exceptions. A company above US$1 million annual gross revenue must use Pro or Mega to own its Assets under those Terms. Upscaling someone else's work does not transfer ownership from its creator.
The user grants Midjourney a broad, perpetual license to submitted Content and generated Assets. Therefore:
  • verify rights to every prompt image, style reference, Omni image, upload, logo, and source asset;
  • obtain documented consent for identifiable people and sensitive portrayals;
  • do not use the service to infringe copyright, patent, or trademark rights;
  • do not promise copyrightability, exclusivity, non-infringement, or trademark clearance; escalate consequential uses to counsel;
  • review third-party likeness, location, cultural, and product rights independently.
根据2026年5月27日生效的服务条款，合规用户在适用法律允许的最大范围内拥有资产所有权，但需受第三方权利和列明的例外情况约束。年总收入超过100万美元的公司必须使用Pro或Mega套餐，才能根据这些条款拥有其资产。放大他人的作品不会将所有权从创作者转移给你。
用户授予Midjourney对提交的内容和生成的资产的广泛、永久许可。因此：
  • 核实每个Prompt图像、风格参考、Omni图像、上传内容、Logo和源资产的权利；
  • 获得可识别个人和敏感描绘的书面授权；
  • 不得使用服务侵犯版权、专利或商标权；
  • 不得承诺可版权性、排他性、不侵权或商标合规；将重大使用需求提交给法务处理；
  • 独立审查第三方肖像权、场所权、文化权和产品权。

Safety and truthful provenance

安全与真实来源

The Community Guidelines apply to public, private, hidden, deleted, Stealth, direct-message, and website work. They require SFW content; prohibit abusive or harmful material, misinformation/disinformation, political-campaign or election-influence images, and deception/fraud; and say not to intentionally mislead recipients about a generated image's nature or source. Passing automated moderation does not prove that a request is allowed.
For every final, keep a production ledger with:
  • creation date, account/plan context, model/version, full prompt and parameters;
  • job ID, seed, visibility mode, input/reference file hashes or stable IDs, and resolved style/profile codes;
  • source, license, and consent for each input;
  • selected output, variations/edits/upscales, and all post-production changes;
  • reviewer, safety/rights disposition, and any AI disclosure required by the channel.
Do not claim Midjourney supplies cryptographic provenance, C2PA credentials, or guaranteed watermarking unless the actual exported file is independently verified to contain it. Add durable provenance or disclosure in the downstream publishing workflow when required.
社区准则适用于公开、私有、隐藏、已删除、Stealth、私信和网站端的所有作品。要求内容为SFW（适合工作场合）；禁止辱骂性或有害材料、虚假信息/误导信息、政治竞选或影响选举的图像，以及欺骗/欺诈行为；并规定不得故意误导接收者关于生成图像的性质或来源。通过自动审核并不代表请求是被允许的。
每个最终作品都需保存生产台账，包含以下内容：
  • 创建日期、账户/套餐背景、模型/版本、完整的Prompt和参数；
  • 任务ID、seed、可见性模式、输入/参考文件的哈希值或稳定ID，以及解析后的风格/配置代码；
  • 每个输入的来源、许可和授权证明；
  • 选中的输出、变体/编辑/放大版本，以及所有后期制作修改；
  • 审核人、安全/权利处理结果，以及发布渠道要求的任何AI披露信息。
除非独立核实实际导出的文件中包含相关内容，否则不得声称Midjourney提供加密来源证明、C2PA凭证或保证水印。必要时在下游发布工作流中添加持久的来源标注或披露信息。

QA gates

QA关卡

Before manual submission

手动提交前

  • The official UI, actual account holder, plan, speed, visibility, and budget are identified.
  • The model supports every reference and parameter combination.
  • References have a documented role and upload rights; real-person consent is recorded.
  • Exact text, logo, product, and geometry needs have a post-production plan.
  • The prompt is SFW, non-deceptive, and consistent with the Terms and applicable law.
  • The projected expansion from repeats/permutations is confirmed.
  • 已明确官方UI、实际账户持有者、套餐、速度模式、可见性和预算。
  • 模型支持所有参考素材和参数的组合。
  • 参考素材有明确的作用和上传权利；真人授权已记录在案。
  • 精确的文字、Logo、产品和几何形状需求已有后期制作方案。
  • Prompt为SFW内容，无欺骗性，符合服务条款和适用法律。
  • 已确认Repeat/排列组合预计扩展的任务数量。

Before approval

审批前

  • Brief: correct subject count, pose/action, composition, crop, palette, mood, and negative space.
  • Continuity: face, hair, wardrobe, props, product form, and distinctive marks checked against authorized references.
  • Image integrity: hands, eyes, teeth, edges, reflections, shadows, occlusion, repeated structures, and impossible geometry inspected at 100%.
  • Text/brand: every glyph and logo replaced or verified; no accidental watermark, pseudo-signature, or competing brand.
  • Pattern: a 
    --tile
    output tested in at least a 3×3 repeat at the final file resolution; seam checked after any upscale.
  • Edit drift: pre/post Editor, Retexture, variation, and upscale compared for lost invariants.
  • Delivery: pixel dimensions, aspect, color profile, alpha/background, compression, file naming, and safe areas meet the destination specification.
  • Governance: visibility, rights, consent, safety, disclosure, and provenance ledger complete.
Reject and repair rather than rationalize a missed hard constraint.
  • 需求符合度： 主体数量、姿势/动作、构图、裁剪、调色板、氛围和留白正确。
  • 连续性： 面部、发型、服装、道具、产品形态和独特标记已与授权参考素材对比核实。
  • 图像完整性： 已100%放大检查手部、眼睛、牙齿、边缘、反射、阴影、遮挡、重复结构和不可能的几何形状。
  • 文字/品牌： 每个字形和Logo都已替换或核实；无意外水印、伪签名或竞争品牌。
  • 图案： 
    --tile
    输出已在最终文件分辨率下进行至少3×3重复测试；任何放大后都已检查接缝。
  • 编辑偏移： 已对比Editor、Retexture、变体生成和放大前后的内容，确认未丢失不变项。
  • 交付要求： 像素尺寸、宽高比、色彩配置文件、alpha/背景、压缩、文件命名和安全区域符合目标规范。
  • 治理合规： 可见性、权利、授权、安全、披露和来源台账完整。
对于未满足的硬性约束，应拒绝并修复，而非找借口合理化。

Common failure diagnoses

常见问题诊断

SymptomLikely causeRepair
Prompt looks attractive but ignores layoutStylize/reference influence too high; too many competing clausesLower 
--s
, 
--sw
, 
--iw
, or 
--ow
; simplify; state spatial relationships; test one influence at a time.
Character/object drifts between scenesWrong reference class, V8.1 selected with Omni intent, or edit crossed into V6.1Use one clean V7 Omni ref; pin 
--v 7
; moderate 
--ow
; minimize current Editor round trips; composite exact details.
Campaign style changed after rerunMutable personalization, unpinned 
--sv
, model default changed, or seed over-trusted		Save resolved codes, pin model/
--sv
, archive refs and prompt, compare against golden outputs.
Aspect remix looks stretchedRemix changed ratio without adding canvasUse Pan, Custom Zoom, or Editor; rebuild composition at target 
--ar
.
Upscaled tile has seamsUpscaler altered edge pixelsUse the pre-upscale tile or repair edges externally; validate the final export in 3×3 repetition.
Text is almost correctGenerative typography limitationReduce phrase length, use double quotes/Raw/lower stylize for concepting, then replace exact copy in post.
“Private” job is visibleStealth absent, used a shared channel, existing job not privatized, or private image published via profile/spotlightVerify Pro/Mega Stealth before creation; use web/DM/private server; privatize existing jobs; audit profile/Explore state.
Old 
--cref
workflow fails in V7		Character Reference is a V6/Niji 6 featureReplace with one Omni image, 
--oref
, and 
--ow
; rebaseline because behavior differs.
Agent proposes a Discord bot or endpointUI concepts were mistaken for an integration contractStop. Return to offline preparation + manual official-UI submission; require explicit provider authorization for any exception.
症状可能原因修复方法
Prompt生成的图很好看但忽略了布局Stylize/参考素材影响过高；冲突条款过多降低 
--s
、
--sw
、
--iw
或 
--ow
；简化Prompt；明确空间关系；每次只测试一种影响因素。
角色/物体在不同场景中偏移参考类型错误、选择了V8.1但想用Omni功能，或编辑时切换到了V6.1使用一张清晰的V7 Omni参考图；固定 
--v 7
；调整 
--ow
至适中；尽量减少当前Editor的往返编辑；合成精确细节。
重新运行后项目风格改变个性化配置可变、未固定 
--sv
、模型默认版本改变，或过度信任seed		保存解析后的代码，固定模型/
--sv
，归档参考素材和Prompt，与标准输出对比。
宽高比Remix后图像拉伸Remix改变了比例但未增加画布使用Pan、Custom Zoom或Editor；在目标 
--ar
下重新生成构图。
放大后的图案有接缝放大工具修改了边缘像素使用放大前的图案或在外部修复边缘；通过3×3重复验证最终导出文件。
文字几乎正确生成式排版的局限性缩短短语长度，使用双引号/Raw模式/降低stylize值进行概念设计，然后在后期替换精确文本。
“私有”任务可见未开启Stealth、使用了共享频道、现有任务未设为私有，或私有图像通过头像/精选内容公开创建前核实Pro/Mega套餐的Stealth权限；使用网页端/私信/私人服务器；将现有任务设为私有；审核头像/Explore页面状态。
旧的 
--cref
工作流在V7中失效		Character Reference是V6/Niji 6的功能替换为一张Omni图像、
--oref
和 
--ow
；由于行为不同，需重新建立基准。
Agent提议使用Discord机器人或端点将UI概念误认为集成协议立即停止。回到离线准备+手动官方UI提交的模式；任何例外情况都需要官方明确授权。

Complete examples

完整示例

Each example is illustrative, not a required formula. Replace subjects, rights context, and parameters to match the brief.
每个示例仅作说明，并非必须遵循的公式。根据需求简报替换主体、权利背景和参数。

Example 1 — controlled product campaign with exact label in post

示例1 — 受控产品宣传，后期添加精确标签

Intent: produce a 4:5 hero concept for a licensed ceramic skincare bottle while preserving a clean label zone.
Interface/model: authorized user submits manually in the website Imagine bar, V8.1 SD for exploration, then HD or approved upscale after layout lock.
Inputs and constraints: one authorized bottle photo assigned as Image Prompt; exact logo and legal copy must not be generated; warm limestone set; one bottle only; label faces camera.
Complete prompt:
text
one matte ivory ceramic skincare bottle standing upright on a pale limestone plinth, cap fully visible, front label panel blank and facing camera, soft late-afternoon window light from camera left, restrained sand and cream palette, premium editorial product photography, eye-level three-quarter view, crisp bottle silhouette, gentle realistic contact shadow, generous clean negative space above and to the right --v 8.1 --sd --ar 4:5 --raw --s 70 --iw 1.2 --chaos 4 --no text, logo, watermark, flowers
Why: Raw and low stylize favor literal layout; moderate image weight carries form without asking for an exact copy; low chaos stabilizes comparisons; the blank panel creates a post-production target.
Expected result: multiple product-photo concepts with one centered bottle, readable silhouette, and a clean label surface.
Likely failures and response: warped cap or changed bottle proportions → reduce style/reference conflict and composite the licensed packshot; pseudo-lettering → repair/paint blank; unwanted crop → regenerate at 4:5 rather than stretch with Remix. Add approved vector label, legal copy, and exact color externally, then inspect reflections and contact shadow.
Meaningful variation: test 
--s {40,70,110}
as a three-job Fast permutation only after the account holder confirms cost.
意图： 为一款授权陶瓷护肤瓶制作4:5的主视觉概念，同时保留干净的标签区域。
界面/模型： 授权用户在网站Imagine栏手动提交，使用V8.1 SD进行探索，布局锁定后再生成HD或获批的放大版本。
输入与约束： 一张授权的瓶子照片指定为图像Prompt；不得生成精确的Logo和法律文本；暖色调石灰岩场景；仅一个瓶子；标签朝向相机。
完整Prompt：
text
one matte ivory ceramic skincare bottle standing upright on a pale limestone plinth, cap fully visible, front label panel blank and facing camera, soft late-afternoon window light from camera left, restrained sand and cream palette, premium editorial product photography, eye-level three-quarter view, crisp bottle silhouette, gentle realistic contact shadow, generous clean negative space above and to the right --v 8.1 --sd --ar 4:5 --raw --s 70 --iw 1.2 --chaos 4 --no text, logo, watermark, flowers
设计思路： Raw模式和低stylize值有利于写实布局；适中的图像权重可保留形态而不会要求精确复制；低chaos值保证对比的稳定性；空白面板为后期制作提供目标区域。
预期结果： 多个产品摄影概念，瓶子居中，轮廓清晰，标签表面干净。
常见问题与应对： 瓶盖变形或瓶子比例改变 → 减少风格/参考素材的冲突，合成授权的产品实拍图；伪文字 → 修复/涂成空白；裁剪不符合预期 → 以4:5比例重新生成，而非用Remix拉伸。在外部添加经批准的矢量标签、法律文本和精确色彩，然后检查反射和接触阴影。
有意义的变体： 仅在账户持有者确认成本后，以三个Fast任务的排列组合测试 
--s {40,70,110}
。

Example 2 — V7 character continuity across a new scene

示例2 — V7角色在新场景中的连续性

Intent: depict an authorized fictional mascot in a rainy transit station while keeping recognizable costume and silhouette.
Interface/model: one clean full-body design sheet assigned as Omni Reference on the website; manual V7 submission in Relax mode on a Standard, Pro, or Mega plan. Verify the live mode before submitting because an account-level Fast default is incompatible with Omni.
Inputs and constraints: user owns the mascot; one Omni image; no region editing until continuity approval; no exact typography.
Complete prompt:
text
the same friendly teal courier mascot waiting beneath a glass transit canopy in light rain, yellow messenger coat zipped halfway, red rectangular satchel across the body, full body visible, three-quarter standing pose, wet pavement reflections, quiet blue-hour city, cinematic illustrated advertising frame, medium-wide eye-level composition, soft rim light, clear silhouette --v 7 --ar 16:9 --s 120 --oref <AUTHORIZED_MASCOT_IMAGE_URL> --ow 220 --chaos 3 --relax --no text, logo, crowd, umbrella
Why: V7 is pinned because Omni requires it; one identity reference has one job; 
--ow 220
is strong but below the typical 400 caution; low chaos prioritizes continuity.
Expected result: scene and pose change while major colors, costume, satchel, and silhouette remain recognizable.
Likely failures and response: emblem/satchel geometry drifts → composite exact brand elements; over-literal reference pose → lower 
--ow
and strengthen pose language; need to fix a hand → remove Omni before current Editor use, make the smallest contextual repair, then compare identity because the edit routes through V6.1.
Meaningful variation: add an authorized Style Reference at 
--sw 80
to move lighting/texture while retaining the same Omni image; test it separately before combining.
意图： 在雨天的公交站中描绘一个授权的虚构吉祥物，保持其服装和轮廓的可识别性。
界面/模型： 一张清晰的全身设计稿在网站端指定为Omni Reference；在Standard、Pro或Mega套餐下，以Relax模式手动提交V7任务。提交前核实当前模式，因为账户级的Fast默认设置与Omni不兼容。
输入与约束： 用户拥有吉祥物版权；一张Omni图像；在连续性获批前不进行区域编辑；无精确排版需求。
完整Prompt：
text
the same friendly teal courier mascot waiting beneath a glass transit canopy in light rain, yellow messenger coat zipped halfway, red rectangular satchel across the body, full body visible, three-quarter standing pose, wet pavement reflections, quiet blue-hour city, cinematic illustrated advertising frame, medium-wide eye-level composition, soft rim light, clear silhouette --v 7 --ar 16:9 --s 120 --oref <AUTHORIZED_MASCOT_IMAGE_URL> --ow 220 --chaos 3 --relax --no text, logo, crowd, umbrella
设计思路： 固定V7版本，因为Omni需要该版本；一个身份参考只承担一个作用；
--ow 220
属于较强水平，但低于通常的400警戒值；低chaos值优先保证连续性。
预期结果： 场景和姿势改变，但主要色彩、服装、挎包和轮廓仍可识别。
常见问题与应对： 徽章/挎包几何形状偏移 → 合成精确的品牌元素；参考姿势过于写实 → 降低 
--ow
并强化姿势描述；需要修复手部 → 在使用当前Editor前移除Omni，进行最小范围的上下文修复，然后对比身份特征，因为编辑通过V6.1处理。
有意义的变体： 添加一个授权的风格参考，设置 
--sw 80
，在保留同一张Omni图像的同时改变布光/纹理；在组合前先单独测试。

Example 3 — seamless textile concept

示例3 — 无缝纺织品概念

Intent: make a repeatable botanical motif for concept approval, not a production-separated print file.
Interface/model: manual V8.1 website or Discord image job.
Inputs and constraints: flat two-color motif, no gradients, no text, seamless export required.
Complete prompt:
text
small hand-cut gingko leaves and round seed pods arranged as a balanced scattered textile motif, flat screen-print shapes, two inks only: deep forest green and warm parchment, consistent element scale, generous breathing room, clean crisp edges, no central focal point --v 8.1 --sd --tile --ar 1:1 --raw --s 90 --chaos 8 --no text, border, frame, gradient, shadow
Why: square tile and 
--tile
establish repeat intent; flat positive constraints are clearer than relying on negatives; modest chaos explores arrangements.
Expected result: a single square pattern unit whose opposite edges are designed to repeat.
Likely failures and response: apparent seam only after upscale → use/test SD output or repair final edges externally; motif collisions in a repeat → choose another variation; unintended third color → separate/recolor in print software.
QA: assemble a 3×3 preview from the exact delivered file, inspect all seams and obvious repetition, then perform color separation and trapping outside Midjourney.
意图： 制作可重复的植物图案用于概念审批，而非生产用的分色印花文件。
界面/模型： 手动在V8.1网站端或Discord提交图像任务。
输入与约束： 平面双色图案，无渐变，无文字，需要无缝导出。
完整Prompt：
text
small hand-cut gingko leaves and round seed pods arranged as a balanced scattered textile motif, flat screen-print shapes, two inks only: deep forest green and warm parchment, consistent element scale, generous breathing room, clean crisp edges, no central focal point --v 8.1 --sd --tile --ar 1:1 --raw --s 90 --chaos 8 --no text, border, frame, gradient, shadow
设计思路： 方形尺寸和 
--tile
参数明确了重复图案的意图；平面化的正向约束比依赖反向描述更清晰；适中的chaos值可探索不同的排列方式。
预期结果： 单个方形图案单元，其对边设计为可重复拼接。
常见问题与应对： 仅放大后出现明显接缝 → 使用/测试SD输出，或在外部修复最终边缘；重复时图案碰撞 → 选择另一个变体；意外出现第三种颜色 → 在印刷软件中分色/重新着色。
QA： 用实际交付的文件组装3×3预览，检查所有接缝和明显的重复痕迹，然后在Midjourney之外进行分色和陷印处理。

Example 4 — poster concept with reliable final typography

示例4 — 最终排版可靠的海报概念

Intent: art-direct a vertical festival poster whose final title must read exactly “NIGHT ORBIT / 18 OCT”.
Interface/model: V8.1 Draft on the website for composition search, then manual SD/HD refinement; final typography in a layout tool.
Draft direction to Conversational mode:
text
Explore 24 vertical poster compositions for an experimental astronomy festival. Use one luminous orbital ring above a dark horizon, cobalt and fluorescent coral, Swiss-modernist spacing, and reserve a clean upper-left title block. Do not render final copy.
Approved refinement prompt:
text
experimental astronomy festival poster background, one luminous orbital ring suspended above a dark geometric horizon, cobalt black field with fluorescent coral glow, precise Swiss-modernist grid, clean upper-left rectangular area reserved for typography, dramatic scale contrast, subtle paper grain, no readable words --v 8.1 --hd --ar 2:3 --raw --s 140 --chaos 2 --no text, letters, logo, watermark, stars
Why: Draft selects hierarchy and palette, not detail; the refinement explicitly reserves a type zone; exact copy is omitted from generation.
Expected result: a typographically useful background with controlled negative space.
Likely failures and response: pseudo-glyphs → remove/repair them before layout; glow enters title block → region repair or choose another candidate; HD edit needed → accept that Editor downshifts to SD and upscale only after approval.
Finish: place the exact title and date as vector text; verify spelling, hierarchy, contrast, safe area, bleed, and print/export profile.
意图： 指导一张竖版节日海报的艺术方向，其最终标题必须精确为“NIGHT ORBIT / 18 OCT”。
界面/模型： 使用网站端V8.1 Draft探索构图，然后手动进行SD/HD优化；最终排版在布局工具中完成。
给对话模式的Draft指令：
text
Explore 24 vertical poster compositions for an experimental astronomy festival. Use one luminous orbital ring above a dark horizon, cobalt and fluorescent coral, Swiss-modernist spacing, and reserve a clean upper-left title block. Do not render final copy.
获批的优化Prompt：
text
experimental astronomy festival poster background, one luminous orbital ring suspended above a dark geometric horizon, cobalt black field with fluorescent coral glow, precise Swiss-modernist grid, clean upper-left rectangular area reserved for typography, dramatic scale contrast, subtle paper grain, no readable words --v 8.1 --hd --ar 2:3 --raw --s 140 --chaos 2 --no text, letters, logo, watermark, stars
设计思路： Draft用于选择层级和调色板，而非细节；优化Prompt明确预留了文字区域；生成时省略了精确文本。
预期结果： 适合排版的背景，留白区域可控。
常见问题与应对： 伪字形 → 排版前移除/修复；光晕进入标题区域 → 区域修复或选择另一个候选方案；需要HD编辑 → 接受Editor会降级为SD的事实，仅在获批后再放大。
收尾： 将精确的标题和日期作为矢量文本放置；核实拼写、层级、对比度、安全区域、出血位和印刷/导出配置文件。

Example 5 — migration and reproducibility audit

示例5 — 迁移与可复现性审核

Intent: update a V6 campaign that used Character Reference and an old style-reference behavior without pretending outputs can be pixel-matched.
Inputs: archived prompts, V6 outputs, source character image, style image/code, 
--cref
, 
--cw
, model/version, seed, and rights records.
Workflow:
  1. Reproduce one control with the original supported V6 settings where available; archive it as the baseline.
  2. Create a V7 branch: replace 
    --cref
    /
    --cw
    with one 
    --oref
    /
    --ow
    , pin 
    --v 7
    , switch to a verified Omni-compatible mode such as Relax on Standard+, and begin near 
    --ow 100–250
    rather than copying 
    --cw
    numerically.
  3. Pin the old Style Reference behavior with 
    --sv 4
    for the first comparison; then test current 
    --sv 6
    as a separate branch.
  4. Keep prompt, aspect ratio, stylize, reference files, and selection rubric constant. Do not use the same seed as proof of equivalence.
  5. Score identity silhouette, costume colors, style palette, composition, anatomy, text/logo drift, and policy/rights compliance. Approve a new golden output and record that migration is perceptual, not deterministic.
Example migrated prompt:
text
one fictional lighthouse keeper matching the authorized reference, walking along a windswept cliff path, navy wool coat, brass lantern in right hand, full figure, graphic gouache illustration, muted storm palette, broad simplified shapes --v 7 --ar 3:2 --s 160 --oref <AUTHORIZED_KEEPER_IMAGE_URL> --ow 180 --sref <AUTHORIZED_STYLE_IMAGE_URL> --sw 100 --sv 4 --chaos 2 --relax --no text, logo
Failure conditions: asserting the seed guarantees the old asset; leaving 
--cref
in V7; changing 
--sv
, model, prompt, and reference simultaneously; or approving without a rights/consent audit.
意图： 更新一个使用了Character Reference和旧版风格参考行为的V6项目，不假装输出可以像素级匹配。
输入： 归档的Prompt、V6输出、源角色图像、风格图像/代码、
--cref
、
--cw
、模型/版本、seed和权利记录。
工作流：
  1. 尽可能使用原始支持的V6设置复现一个对照组；将其归档作为基准。
  2. 创建V7分支：将 
    --cref
    /
    --cw
    替换为一个 
    --oref
    /
    --ow
    ，固定 
    --v 7
    ，切换到经验证的Omni兼容模式（例如Standard+套餐的Relax模式），初始 
    --ow
    值设在100-250附近，而非直接复制 
    --cw
    的数值。
  3. 第一次对比时，用 
    --sv 4
    固定旧版风格参考行为；然后作为单独分支测试当前的 
    --sv 6
    。
  4. 保持Prompt、宽高比、stylize、参考文件和筛选标准不变。不要使用相同的seed作为等效性证明。
  5. 对身份轮廓、服装颜色、风格调色板、构图、人体结构、文字/Logo偏移和政策/权利合规性进行评分。批准新的标准输出，并记录迁移是感知层面的，而非确定性的。
迁移后示例Prompt：
text
one fictional lighthouse keeper matching the authorized reference, walking along a windswept cliff path, navy wool coat, brass lantern in right hand, full figure, graphic gouache illustration, muted storm palette, broad simplified shapes --v 7 --ar 3:2 --s 160 --oref <AUTHORIZED_KEEPER_IMAGE_URL> --ow 180 --sref <AUTHORIZED_STYLE_IMAGE_URL> --sw 100 --sv 4 --chaos 2 --relax --no text, logo
失败条件： 声称seed能保证复现旧资产；在V7中保留 
--cref
；同时改变 
--sv
、模型、Prompt和参考素材；或未经权利/授权审核就批准。

First-party source ledger

第一方来源台账

All links below were checked 2026-07-09. Treat dates, defaults, prices, availability, limits, and policy language as volatile and re-check before consequential work.
以下所有链接已于2026年7月9日核查。日期、默认设置、价格、可用性、限制和政策条款均可能变动，在进行重要工作前请重新核实。