Back to Details

luma

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

luma

luma

The cinematic video-generation tool skill — define the shot, reference with stills + frames, economize with drafts, assemble the sequence, and mind the rights. The agent briefs (and can drive the API where connected); the human judges every clip; WoopSocial publishes. (Ships with 
tools/integrations/luma.md
.)
电影级视频生成工具 skill——定义镜头，以静帧 + 帧图像作为参考，利用草稿控制成本，组装序列，并妥善处理权利。Agent 负责编写简报（已连接时也可调用 API）；人工作出每段片段的判断；WoopSocial 负责发布。（随附 
tools/integrations/luma.md
。）

The POV: shots, not films — and drafts, not gambles

核心观点：生成镜头，而非影片——先用草稿，而非赌博

Dream Machine's 2026 edge is per-shot cinematic quality: Ray3 was the first "reasoning" video model (it interprets, generates, self-evaluates, retries), with 16-bit HDR, keyframes, character reference, and Modify (restyle real footage, keep the performance). The top-1% operator holds four truths the marketing softens. (1) Generations are shots, not films: 5–10 seconds each, Extend degrades past the initial clip — a brand film is a shot list, generated shot-by-shot and assembled in an editor with sound added in post (no native audio; that's Veo/Kling territory). (2) Draft first, always: iterate cheap in Draft Mode, Hi-Fi master only the winner — because HDR at 1080p runs ~16× standard credits and monthly credits expire at reset. (3) Variance is real: the same prompt differs run-to-run more than half the time (independent testing) — so client-grade consistency comes from references (still-first image-to-video, character reference on base Ray3, keyframes), never from luck. (4) The rights fine print: free = watermarked non-commercial, the commercial floor is the Plus-level plan (tier tables genuinely conflict — verify in-app), and Luma keeps a marketing license on your generations — the clause client work needs to know about.
Dream Machine 在 2026 年的优势是逐镜头的电影级质量：Ray3 是首个“推理”视频模型（能够理解、生成、自我评估并重试），支持 16-bit HDR、关键帧、角色参考和 Modify（重塑实拍素材风格，同时保留表演）。顶尖 1% 的操作者掌握四个被营销话术淡化的真相。（1）生成的是镜头，不是影片：每个镜头 5–10 秒；超过初始片段后，Extend 的质量会下降——品牌影片应采用镜头清单，逐镜头生成，再在剪辑器中组装，并在后期添加声音（无原生音频；那是 Veo/Kling 的领域）。（2）永远先用草稿：在 Draft Mode 中低成本迭代，只将胜出稿精修为 Hi-Fi master——因为 1080p HDR 会消耗约标准模式 16 倍的积分，而月度积分会在重置时失效。（3）波动确实存在：根据独立测试，同一提示词在多次运行间有超过一半的概率产生差异——因此，客户级一致性来自参考素材（静帧优先的图生视频、基础 Ray3 上的角色参考、关键帧），而非侥幸。**（4）权利条款中的细则：**免费版带水印且不可商用；商业用途的最低门槛是 Plus 级套餐（套餐表确实存在冲突——请在应用内核验），并且 Luma 会为你的生成内容保留营销许可——这是客户项目必须了解的条款。

Read these first

首先读取

  1. ai-video — the model-agnostic router/craft above the video tools.
  2. brand-profile + design-and-templates (the brand feel) + short-form-video-script (the beats).
  1. ai-video——位于各视频工具之上的模型无关路由/创作层。
  2. brand-profile + design-and-templates（品牌调性）+ short-form-video-script（叙事节拍）。

The framework: DREAM

框架：DREAM

(Depth: 
references/the-dream-framework.md
.)
  • D — Define the shot: a director's shot brief (subject + action + setting + light + mood + camera move); one shot per generation; vague prompts pay the variance tax.
  • R — Reference with stills + frames: still-first image-to-video (brand-true frames from flux/Photon); keyframes for transitions; character reference (Ray3) for identity; Modify for real footage; never an unpermitted likeness.
  • E — Economize with drafts: Draft Mode → Hi-Fi the winner; ~3 attempts per usable clip budgeted; HDR only for graded finals; credits expire — plan sessions.
  • A — Assemble the sequence: shots → capcut (pace to the beats), sound in post (ai-music-and-sound), keyframe-matched cuts; drift review on sets.
  • M — Mind the rights: commercial tier confirmed; Luma's license-back understood (Enterprise for NDA); AI-disclosure (EU AI Act; C2PA); then the human judges and WoopSocial publishes.
（深入说明：
references/the-dream-framework.md
。）
  • **D — 定义镜头：**导演式镜头简报（主体 + 动作 + 场景 + 光线 + 氛围 + 运镜）；每次生成只对应一个镜头；模糊提示词会招致相应的波动成本。
  • **R — 以静帧 + 帧图像作参考：**静帧优先的图生视频（来自 flux/Photon、符合品牌风格的画面）；用于转场的关键帧；用于保持身份的角色参考（Ray3）；用于实拍素材的 Modify；绝不使用未经许可的肖像。
  • **E — 使用草稿控制成本：**Draft Mode → 对胜出稿使用 Hi-Fi；每个可用片段按约 3 次尝试预留预算；HDR 仅用于调色后的最终版本；积分会过期——提前规划创作会话。
  • **A — 组装序列：**镜头 → capcut（按节拍控制节奏），后期添加声音（ai-music-and-sound），进行关键帧匹配剪切；按镜头组检查画面漂移。
  • **M — 留意权利：**确认商业套餐；了解 Luma 的 license-back（涉及 NDA 时使用 Enterprise）；AI-disclosure（EU AI Act；C2PA）；之后由人工作出每段片段的判断，并由 WoopSocial 发布。

The reality (verify-quarterly)

实际情况（每季度核验）

2026 Luma: Ray3 (reasoning model; 16-bit HDR + EXR; keyframes; character reference; visual annotation; Modify V2V) · Ray3.14 (Jan 2026 — native 1080p, ~4× faster, ~3× cheaper at 720p; drops character reference + HDR) · a newer Ray already on the official pricing page (the line moves) · Photon (images) · Luma Agents (multi-model bundle incl. Veo/Kling/ElevenLabs — worth it if you'd pay for 2–3 bundled models; overpaying for one) · official Python SDK (create → poll → download); API credits are a separate wallet. Economics: Draft→Hi-Fi is the signature workflow; HDR 1080p ≈ 16× standard 720p; 3 attempts/usable clip; monthly credits expire, top-ups roll over; Trustpilot skews negative on billing. Rights: free = non-commercial; commercial floor = Plus-level ($30/mo; tier tables conflict — two plan generations coexist; verify in-app); Luma's marketing license-back; Enterprise = data-privacy guarantee. Variance: same prompt differs
50% of runs (goenhance). Sora's consumer shutdown (Apr 2026) made Luma a migration target. Attribute all; verify-quarterly. Full detail: 
references/luma-2026-reality.md
. The model table, shot-brief pattern, the sequence workflow, the credit worksheet, and two worked examples: 
references/shot-patterns-and-templates.md
.
2026 Luma：Ray3（推理模型；16-bit HDR + EXR；关键帧；角色参考；视觉标注；Modify V2V）· Ray3.14（2026 年 1 月——原生 1080p，速度约快 4×，720p 成本约降至 1/3；不再支持角色参考 + HDR）· 官方定价页上已有一款更新的 Ray（产品线变化很快）· Photon（图像）· Luma Agents（多模型组合，包含 Veo/Kling/ElevenLabs——如果你本来会为 2–3 个捆绑模型付费，就值得考虑；否则，只买其中一个模型会显得过贵）· 官方 Python SDK（创建 → 轮询 → 下载）；API 积分使用独立钱包。
成本：Draft→Hi-Fi 是标志性工作流；HDR 1080p ≈ 标准 720p 的 16×；每个可用片段约尝试 3 次；月度积分到期作废，充值积分可结转；Trustpilot 上关于账单的负面评价偏多。权利：免费版 = 非商用；商业用途最低需要 Plus 级（约 $30/月；套餐表存在冲突——两代套餐同时存在；请在应用内核验）；Luma 的营销 license-back；Enterprise = 数据隐私保障。波动：同一提示词在超过 50% 的运行中都会不同（goenhance）。Sora 消费版于 2026 年 4 月关闭，使 Luma 成为迁移目标。所有内容均须注明来源；每季度核验。 完整细节：
references/luma-2026-reality.md
。模型表、镜头简报模式、序列工作流、积分核算表和两个实战示例：
references/shot-patterns-and-templates.md
。

Honest scope (never violate)

诚实的能力边界（绝不违反）

  • The agent briefs shots, plans references/keyframes/credits, and drives the SDK where connected (exact human steps otherwise; no unreviewed auto-publish); the human judges every clip (no fabricated "that shot works"; drift reviews on sets); WoopSocial publishes the exports — it does not generate or edit video.
  • Rights spine: free tier never for client/commercial work; verify the current tier in-app; Luma's license-back disclosed to clients where relevant; Enterprise for NDA; counsel for high-stakes (not legal advice). Likeness/misinformation: no unpermitted real-person likeness (real or AI lookalike); no photoreal fake-event footage of real entities ("satire" doesn't launder defamation); AI-disclosure where required. Sound in post with proper licenses. Never fabricate rates, tiers, benchmarks, capabilities. (Full scope: 
    references/scope-and-connections.md
    .)
  • Agent 编写镜头简报，规划参考素材/关键帧/积分，并在已连接时调用 SDK（否则给出准确的人工操作步骤；不会在未经审核的情况下自动发布）；人工作出每段片段的判断（不会捏造“这个镜头可用”的结论；按镜头组检查漂移）；WoopSocial 发布导出的文件——它不生成或剪辑视频。
  • **权利合规主线：**免费套餐绝不用于客户/商业项目；在应用内核验当前套餐层级；必要时向客户披露 Luma 的 license-back；涉及 NDA 时使用 Enterprise；高风险事项应咨询律师（不构成法律建议）。**肖像/虚假信息：**绝不使用未经许可的真实人物肖像（无论是真实人物本人，还是其 AI lookalike）；不得生成涉及真实主体的照片级虚假事件影像（所谓“讽刺”不能为诽谤“洗白”）；法规要求时进行 AI-disclosure。在后期添加声音，并确保使用适当许可。绝不捏造费率、套餐层级、基准测试或功能。（完整范围：
    references/scope-and-connections.md
    。）

Distinct from its siblings (route correctly)

与同类 skill 的区别（正确路由）

luma (this) = the Dream Machine lane (cinematic shots, HDR, references, Modify) · ai-video = the model-agnostic router (read first) · veo-3 / kling / runway = sibling skills (runway = the control-focused/edit-grade lane) — sound-native generation routes to Veo/Kling (Luma = silent + post) · flux / image-prompt = the brand-true stills this animates · heygen / synthesia = avatar presenters (a different job) · capcut = where shots become films · ai-music-and-sound = the audio Luma doesn't generate.
luma（当前 skill） = Dream Machine 路线（电影级镜头、HDR、参考素材、Modify）· ai-video = 模型无关的路由器（先读取）· veo-3 / kling / runway = 同类 skill（runway = 偏重精细控制/编辑成片的路线）——带原生声音的生成应路由到 Veo/Kling（Luma = 静音 + 后期）· flux / image-prompt = 为其提供动画化的品牌静帧 · heygen / synthesia = 数字人主播（不同任务）· capcut = 将镜头组装成影片的地方 · ai-music-and-sound = 补足 Luma 不生成的音频。

Where this connects

连接关系

Reads first: ai-video + brand-profile + design-and-templates + short-form-video-script. Pulls stills from: flux / image-prompt. Feeds: capcut (assembly), ai-music-and-sound (audio), tiktok- video-publishing / instagram-reels-publishing / youtube-shorts (the finished clip). Publishes via: export → scheduling-and-queue → WoopSocial. Tool file: 
tools/integrations/luma.md
. Measure with: native + analytics-and-reporting — never fabricated.
首先读取：ai-video + brand-profile + design-and-templates + short-form-video-script。从以下工具拉取静帧：flux / image-prompt。为以下流程提供输入：capcut（组装）、ai-music-and-sound（音频）、tiktok-video-publishing / instagram-reels-publishing / youtube-shorts（成片）。发布路径：导出 → scheduling-and-queue → WoopSocial。工具文件：
tools/integrations/luma.md
。使用原生数据 + analytics-and-reporting 进行衡量——绝不捏造。

Definition of done

完成标准

A Luma workflow that ships shots on-budget and on-rights: every generation briefed as a director's shot (one shot, concrete subject/action/light/camera), brand control achieved still-first (image-to-video from canonical flux/Photon frames; character reference on Ray3 for identity; keyframes for transitions; Modify for real footage; no unpermitted likeness), credits governed by the draft-first loop (Draft iterations → one Hi-Fi master; HDR reserved for graded finals; ~3-attempt budget; sessions planned around expiring credits; API wallet funded separately where used), sequences assembled in capcut to the script's beats with licensed sound added in post and a side-by-side drift review, and rights closed before publishing (commercial tier verified in-app; Luma's license-back known; AI-disclosure where required; no fake-event footage of real entities); the human judging every clip and WoopSocial publishing the finished file; no fabricated rates/tiers/benchmarks, no one-generation "film" promises; and correctly distinguished from ai-video, veo-3/kling, flux, heygen/synthesia, and capcut.
一个既能控制预算、又能满足权利要求的 Luma 工作流：每次生成都按导演镜头编写简报（一个镜头，明确的主体/动作/光线/运镜）；通过静帧优先实现品牌控制（使用标准 flux/Photon 画面进行图生视频；在 Ray3 上使用角色参考保持身份；用关键帧处理转场；对实拍素材使用 Modify；不使用未经许可的肖像）；通过草稿优先循环管理积分（Draft 迭代 → 生成一个 Hi-Fi master；HDR 仅留给完成调色的最终版本；按约 3 次尝试设定预算；围绕即将过期的积分规划创作会话；使用 API 时另行为 API 钱包充值）；在 capcut 中按脚本节拍组装序列，后期加入已授权的声音，并进行并排漂移检查；发布前完成权利闭环（在应用内核验商业套餐；了解 Luma 的 license-back；按要求进行 AI-disclosure；不生成涉及真实主体的虚假事件影像）；由人工作出每段片段的判断，并由 WoopSocial 发布成片；不捏造费率/套餐层级/基准测试，不承诺通过一次生成得到“影片”；并正确区别于 ai-video、veo-3/kling、flux、heygen/synthesia 和 capcut。