AI 视频生成 · 人才地图 · 美 / 英 / 以 / 中亚 / 加AI Video Generation · Talent Map · US / UK / Israel / Central Asia / Canada
七家 AI 视频公司，1,687 名在册成员，Seven AI video companies, 1,687 listed members — 真正做研发的只有 518 人——来自大厂与自动驾驶，影视特效不到 1%。but only 518 are actually building. They come from Big Tech and self-driving — film VFX is under 1%.
拆开 Runway / Luma / Pika / Higgsfield / Decart / HeyGen / Synthesia 的真实研发团队：有多大、谁组成、从哪来、在哪。We break down the real R&D teams at Runway / Luma / Pika / Higgsfield / Decart / HeyGen / Synthesia: how big, who they are, where they come from, where they sit.
报告日期Report date2026-06-22出品Produced byMetix AI覆盖Coverage7 家公司 · 全球7 companies · global
Executive Summary
01核心结论Key takeaways
Metix AI 全球人才库，数据截至 2026 年年中。统计七家公司的在册成员，筛出在职的"工程 + 研究 + 创意设计"核心（下称Metix AI global talent pool, data through mid-2026. We count each of the seven companies' listed members and isolate the current "engineering + research + creative design" core (the 建造者builders）。均为聚合数字，不含个人信息。). All figures are aggregate, with no personal information.
在册 1,687 人，建造者只有 518（三成）。其余是创作者社区与销售。对标团队规模，先把这三层分开——消费端产品尤其要警惕"社区"虚高。Of 1,687 listed members, only 518 are builders (about a third). The rest are creator communities and sales. To benchmark team size, separate these three layers first — and watch out for inflated "community" counts, especially on consumer products.
②结构各异：研究公司 vs 产品公司Different structures: research shops vs. product shops
整体工程 60%、研究 21%、创意 13%。但 Luma 研究占 55%、Higgsfield 工程占 80%——"自研模型"与"应用层产品"对应完全不同的招聘画像。Across the field it's 60% engineering, 21% research, 13% creative. But Luma is 55% research and Higgsfield is 80% engineering — "build-your-own-model" and "application-layer product" map to completely different hiring profiles.
③大厂为主，影视特效极少Big Tech dominates, film VFX barely registers
最大来源是 Google / Meta / Amazon。影视特效背景不到 1%，"好莱坞转入"是个例。真正的视频人才来自机器学习、大厂工程，以及自动驾驶、军工情报、搜索引擎。The biggest feeders are Google / Meta / Amazon. Film-VFX backgrounds are under 1%, and "Hollywood crossover" is the exception. The real video talent comes from machine learning, Big Tech engineering, and self-driving, defense intelligence, and search engines.
④五个人才地理，都在扩张Five talent geographies, all expanding
美实验室、以色列、哈萨克斯坦、英欧、美加——五套互不相通的市场。70% 近两年入职、应届仅 6.8%：纯横向挖角。US labs, Israel, Kazakhstan, UK-Europe, US-Canada — five disconnected markets. 70% joined in the past two years and only 6.8% are new grads: this is pure lateral poaching.
口径。Methodology. 仅统计七家纯视频 / 生成媒体公司；OpenAI Sora、Google Veo 嵌在母公司内、无法单独识别，仅作人才来源与背景，不计入。数字均经独立复核。同款透视可按需为任意目标公司生成。We count only the seven pure-play video / generative-media companies; OpenAI Sora and Google Veo sit inside their parent companies and can't be isolated, so they appear only as feeders and background, not in the totals. All figures are independently verified. The same x-ray can be generated on demand for any target company.
Team Composition
02研发核心有多大、由谁组成How big the R&D core is, and who's in it
在册人数不等于研发人数。建造者只占约三成——下图是各家建造者绝对数与在册总数的对照。Listed headcount isn't R&D headcount. Builders are only about a third — the chart below compares each company's absolute builder count against its total listed members.
差额分两类：消费端是The gap splits two ways: on the consumer side it's 创作者社区creator communities，企业端是, and on the enterprise side it's 销售sales。都是真实业务，但都不是研发。. Both are real business functions, but neither is R&D.
公司Company
在册Listed
建造者builders
创作者社区creator communities
销售 / 市场Sales / marketing
建造者占比Builder share
Pika
68
17
27
6
25%
Higgsfield
188
50
60
35
27%
Runway
275
87
60
36
32%
HeyGen
242
93
39
45
38%
Luma AI
200
83
34
42
42%
Decart
91
38
8
19
42%
Synthesia
623
150
27
265
24%
合计Total
1,687
518
255
448
31%
"创作者社区"是谁？Who are the "creator communities"?消费端产品的 Creative Partner / Ambassador 项目成员。他们常把多家产品同时列为"当前身份"（我们见过单人挂十余家），是生态推广者、不是员工——Pika、Higgsfield 的在册数里这一类占三到四成。对标团队时应单列。These are members of consumer products' Creative Partner / Ambassador programs. They often list several products at once as their "current role" (we've seen one person tagged to more than ten), and they're ecosystem promoters, not employees — at Pika and Higgsfield this group makes up 30–40% of the listed count. Break them out separately when benchmarking a team.
建造者由谁组成What the builders are made of
按职能拆开，一眼看出研究驱动还是产品驱动：Luma 研究过半，Higgsfield 工程八成。创意 / 设计各家都有（全行业 13%），但几乎不来自影视特效。Split by function, you can tell at a glance whether a company is research-driven or product-driven: Luma is more than half research, Higgsfield is 80% engineering. Creative / design appears everywhere (13% across the field), but almost none of it comes from film VFX.
各家建造者按职能占比（合计 100%）。数据来源 Metix AIEach company's builders by function (totals to 100%). Source: Metix AI
Where The Talent Comes From
03人才从哪来：大厂为主，影视特效极少Where the talent comes from: Big Tech leads, film VFX barely shows up
建造者的前雇主去重排序，前 11 大来源——大厂占主导。Builders' prior employers, deduplicated and ranked — the top 11 feeders, dominated by Big Tech.
Google
39
Meta
29
Amazon
28
Microsoft
20
Snap
17
Apple
16
字节跳动 ByteDanceByteDance
15
NVIDIA
12
Adobe
11
华为 HuaweiHuawei
10
Yandex
8
柱长 = 曾任职该公司的建造者数（去重 / 人，Google 39 为 100%）。数据来源 Metix AIBar length = number of builders who once worked there (deduplicated, by person; Google's 39 = 100%). Source: Metix AI
只看"视频相关"来源：自动驾驶 3D 视觉、军工情报、搜索引擎排在前面，Looking only at "video-relevant" feeders, self-driving 3D vision, defense intelligence, and search engines come out on top, while 影视特效几乎为零film VFX is essentially zero。.
"自动驾驶"= Waymo / Cruise / Zoox 等；"VFX"= Pixar / ILM / Weta / DNEG 等合计。数据来源 Metix AI"Self-driving" = Waymo / Cruise / Zoox and the like; "VFX" = Pixar / ILM / Weta / DNEG and others combined. Source: Metix AI
VFX 转行？VFX career-switchers?用最宽的口径筛，518 名建造者里影视特效背景Even on the broadest definition, among the 518 builders, film-VFX backgrounds are 不到 1%under 1%，且只零星出现在 Runway / Luma。研发核心由机器学习与大厂工程构成，创意 / 设计也多来自产品设计，而非影视流水线。, and they surface only sporadically at Runway / Luma. The R&D core is built from machine learning and Big Tech engineering, and even the creative / design hires come mostly from product design rather than the film pipeline.
注意。Note. "以色列军工情报 ≈24"几乎全在 Decart 一家（38 人里 24 人）——这是"Decart 是以色列背景公司"，而非赛道普遍现象；Yandex 同理集中在 Higgsfield。详见下一节。The "Israeli defense intelligence ≈24" is almost entirely concentrated in one company, Decart (24 of its 38 builders) — that reflects "Decart is an Israeli-rooted company," not a sector-wide pattern; likewise, Yandex is concentrated at Higgsfield. See the next section for details.
Five Talent Geographies
04一个赛道，五个互不相通的市场One sector, five disconnected markets
按所在地拆开，七家根本不在同一个市场招人。下图是各家的地理分布，卡片给出人才底色与挖人建议。Split by location, the seven companies simply aren't hiring in the same market. The chart below shows each company's geographic distribution, and the cards give the underlying talent profile and poaching advice.
各家建造者所在地占比（合计 100%）。数据来源 Metix AIEach company's builders by location (totals to 100%). Source: Metix AI
① 美国实验室 · Luma / Runway① US labs · Luma / Runway
美国 · 电影级生成US · cinematic generation
83 / 87建造者builders
26.5% · 12.6%博士率PhD rate
71% · 62%在美国in the US
模型核心。来源是 Meta / Google / NVIDIA，加一条暗管——自动驾驶 3D 视觉（Luma 7 人出自 Waymo / Cruise / NVIDIA）。Luma 研究岗 55%、博士 26.5%，最偏研究。The model core. Feeders are Meta / Google / NVIDIA, plus a hidden pipeline — self-driving 3D vision (7 of Luma's people came from Waymo / Cruise / NVIDIA). Luma is 55% research and 26.5% PhDs, the most research-leaning of all.
② 以色列 · Decart② Israel · Decart
特拉维夫 · 实时生成Tel Aviv · real-time generation
38建造者builders
63%军工情报背景defense-intelligence background
76%在以色列in Israel
最精瘦的团队：在册 91 人仅 8 个社区名头。38 名建造者 24 人有 Unit 8200 / IDF 背景，叠加 Technion。典型以色列深科技画像。The leanest team: 91 listed members, only 8 of them community tags. 24 of its 38 builders have Unit 8200 / IDF backgrounds, layered with Technion. A textbook Israeli deep-tech profile.
③ 哈萨克斯坦 · Higgsfield③ Kazakhstan · Higgsfield
阿拉木图 · 增长型产品Almaty · growth product
50建造者builders
80%工程岗engineering roles
68%在哈萨克斯坦in Kazakhstan
团队在哈萨克斯坦、不在硅谷。来源是 Yandex + 本地孵化器（nFactorial、Aviata）。工程驱动、最年轻（任职中位 11 个月），搜索 / 增长底色。The team is in Kazakhstan, not Silicon Valley. Feeders are Yandex plus local incubators (nFactorial, Aviata). Engineering-driven and the youngest of the group (median tenure 11 months), with a search / growth foundation.
④ 美国 / 加拿大 · HeyGen④ US / Canada · HeyGen
旧金山 + 多伦多 · 数字人San Francisco + Toronto · digital humans
工程核心在美加、不在中国。但有清晰的中国大厂管道：前 Snap 9 + 字节 6 + 华为 4。产品工程形态，博士率 3.2%。The engineering core sits in the US and Canada, not China. But there's a clear China Big-Tech pipeline: ex-Snap 9 + ByteDance 6 + Huawei 4. A product-engineering shape, with a 3.2% PhD rate.
⑤ 英国 / 欧洲 · Synthesia⑤ UK / Europe · Synthesia
伦敦 · 企业级数字人London · enterprise digital humans
150建造者builders
23 个月23 months任职中位median tenure
92%在英 / 欧in the UK / Europe
资历最深、最像 B2B SaaS。来源是 Google / Amazon / Microsoft + UCL，92% 在英欧。在册数大是因为 265 人的销售团队，不是社区。The most senior and the most B2B-SaaS-like. Feeders are Google / Amazon / Microsoft plus UCL, with 92% in the UK and Europe. Its large listed count comes from a 265-person sales team, not a community.
＋ 旧金山小队 · Pika＋ San Francisco squad · Pika
旧金山 · 样本偏小San Francisco · small sample
17建造者builders
35%创意 / 设计Creative / design
76%在美国in the US
极小精英团队（17 人，占比仅供参考）。创意 / 设计占比七家最高，来源含 Google 与字节，创始有斯坦福 / Meta AI 背景。A tiny elite team (17 people; percentages are indicative only). The highest creative / design share of the seven, with feeders including Google and ByteDance, and founders out of Stanford / Meta AI.
Credentials & Hiring Pace
05学历与招聘节奏Education and hiring cadence
学历结构印证研究 vs 产品的分野：Luma 博士最多，Higgsfield 以学士为主。（"未填写"= 公开资料无学历信息，不代表无学位。）The education mix confirms the research-vs-product divide: Luma has the most PhDs, Higgsfield is mostly bachelor's degrees. ("Not listed" = no education info in public records, which doesn't mean no degree.)
各家建造者学历占比（合计 100%；Pika 样本小未列）。数据来源 Metix AIEach company's builders by education (totals to 100%; Pika omitted for small sample). Source: Metix AI
招聘节奏：70% 近两年入职Hiring cadence: 70% joined in the past two years，应届仅 6.8%、职业中位 7–14 年——横向挖角为主。下图为各家近 12 个月入职率。, only 6.8% new grads, and a median career length of 7–14 years — overwhelmingly lateral poaching. The chart below shows each company's hires in the past 12 months.
Higgsfield
51%
Luma AI
45%
Decart
40%
HeyGen
38%
Runway
31%
Pika 样本小small sample
31%
Synthesia
26%
近 12 个月入职的建造者占比（Higgsfield 51% 为 100%）。数据来源 Metix AIShare of builders who joined in the past 12 months (Higgsfield's 51% = 100%). Source: Metix AI
外部锚点。External benchmark. 建造者博士率（有学位者）13.7%，与 OpenAI（14.4%）、Anthropic（13.7%）同一量级。拉低整体在册博士率的是社区与销售，不是低水平工程。The builder PhD rate (among those with a degree) is 13.7%, on par with OpenAI (14.4%) and Anthropic (13.7%). What drags down the overall listed PhD rate is community and sales, not low-caliber engineering.
中国管道在来源、不在所在地。The China pipeline is in the feeders, not the locations. 只看雇主：字节跳动 15 + 华为 10 + 腾讯，主要流向 HeyGen 与 Luma。叠加消费端公开的华人创始，中国背景在本赛道的参与度高于自动驾驶 / 芯片。By employer alone: ByteDance 15 + Huawei 10 + Tencent, flowing mainly to HeyGen and Luma. Add the publicly known Chinese founders on the consumer side, and Chinese backgrounds participate more heavily in this sector than in self-driving / chips.
How To Use This
06给 HR / 猎头 / VC 的用法How HR / recruiters / VCs can use this
HR：按研发口径对标HR: benchmark on an R&D basis
对标团队规模，先剥掉创作者社区与销售——竞品"百人团队"的真实研发可能只有三成。画像分型：电影级生成挖大厂 + 自动驾驶 CV，企业数字人挖 B2B SaaS。When benchmarking team size, strip out creator communities and sales first — a competitor's "hundred-person team" may have real R&D of only a third. By profile: for cinematic generation, poach Big Tech plus self-driving CV; for enterprise digital humans, poach B2B SaaS.
猎头：跟着来源走Recruiters: follow the feeders
几乎不招应届，70% 近两年入职——大量人才正处可挖窗口。最大来源是大厂；被低估的是自动驾驶 3D 视觉，以及区域性的以色列 8200、Yandex。Almost no new grads, and 70% joined in the past two years — a large pool of talent is in its poaching window right now. The biggest feeders are Big Tech; the underrated ones are self-driving 3D vision, and regionally, Israel's Unit 8200 and Yandex.
VC：用人才结构做尽调VCs: use talent structure for diligence
职能比例、博士率、来源，直接读出"自研模型"还是"应用层"。"在册人数"要拆三层——掺了多少社区与销售，决定真实研发承载力。Function mix, PhD rate, and feeders read straight off whether a company is "build-your-own-model" or "application-layer." The "listed headcount" must be split into three layers — how much community and sales is mixed in determines the real R&D capacity.
想要某一家的完整名单，或换成你的目标公司？Want the full list for one company, or your own target company instead?
人群清洗、职能构成、人才来源与地理，都可按需为任意一家 AI 视频 / 生成媒体公司单独生成，并对接可联系的候选人。Population cleanup, function mix, feeders, and geography can all be generated on demand for any single AI video / generative-media company, and connected to reachable candidates.
聚合报告 · 不展示任何个人信息 · 由 Metix AI · Mira 提供Aggregate report · no personal information shown · provided by Metix AI · Mira
说明：本报告基于 Metix AI 全球人才库，统计七家 AI 视频 / 生成媒体公司的在册成员（取在职者，再按工程 + 研究 + 创意设计岗位筛出"建造者"），数据截至约 2026 年年中。各项均为可见样本的聚合结果，仅供参考；样本较小的公司（如 Pika 仅 17 名建造者）需谨慎解读；报告不展示任何个人姓名、联系方式或敏感属性。融资 / 估值等公开信息仅作背景。Note: This report is based on the Metix AI global talent pool, counting the listed members of seven AI video / generative-media companies (taking current members, then isolating "builders" by engineering + research + creative design roles), with data through roughly mid-2026. Every figure is an aggregate result over the visible sample and is indicative only; companies with smaller samples (such as Pika, with just 17 builders) should be read with caution; the report shows no individual names, contact details, or sensitive attributes. Funding / valuation and other public information is used only as background.
Metix AI · Mira | AI 视频生成人才地图 2026 | 2026-06-22Metix AI · Mira | AI Video Generation Talent Map 2026 | 2026-06-22Talent analytics powered by Metix AI
