Floweriver01
相机可控生成Camera controlPose
把输入锚定在三维空间：逐帧精确的相机位姿、关键帧锚定与轨迹插值。你坐在导演椅上，而不是拉动老虎机。Inputs anchored in 3D space — frame-accurate camera pose, keyframe anchoring, trajectory interpolation. You sit in the director's chair.
Floweriver · Spatial Intelligence
空间智能，
从世界模型开始
Spatial intelligence,
starting from a world model
流花的空间智能世界模型，把真实世界重建为可控制、可推演、可生成的三维场景——像拍摄一样生成，像物理一样稳定。 Floweriver's world model reconstructs reality into 3D scenes you can control, simulate and generate — generated like a camera move, stable like physics.
01 — Capability
一个模型，四项核心能力 One model. Four capabilities.
不是四个各自为战的模型，而是同一个世界模型在四个方向上的统一表达。每一项能力，都可以直接进入生产流程。 Not four separate models — one world model expressed along four axes. Each capability plugs straight into production.
02
空间重建ReconstructionRebuild
看得越多，想象得越少。从单目视频到点云、网格与 3D 高斯泼溅的统一表达，自由视点实时漫游。The more we see, the less we guess — mono video into point clouds, meshes and Gaussian splats, roamable from any viewpoint.
03
时空模拟Space-timeSimulate
从普通相机到「子弹时间」。时间成为可编辑的维度，为具身智能与自动驾驶规模化生产训练场景。From an ordinary camera to bullet time — time becomes an editable dimension, producing training scenarios at scale.
04
世界生成World generationGenerate
不止生成画面，生成一个可以走进去、可以被编辑的世界——长时序、长距离的空间一致性。Not frames but a world you can walk into and edit — spatially consistent over long horizons and long distances.
02 — Technology
多模态自回归扩散 Transformer A multimodal autoregressive diffusion transformer
统一建模图像、视频、深度与相机参数，让「看」与「造」共享同一套表征。镜头语言是一等公民，而不是后处理。 Images, video, depth and camera parameters share one representation — seeing and making become one. Cinematography is first-class, not post-processing.
数据飞轮 · Real → Sim → RealThe data flywheel · Real → Sim → Real
Architecture
多模态自回归扩散 TransformerMultimodal autoregressive diffusion transformer
视觉 token 与几何 token 共享同一注意力空间Visual and geometric tokens share one attention space
Control
相机位姿 · 关键帧 · 图文条件Camera pose · keyframes · text & image
镜头参数直接进入生成回路Lens parameters enter the generation loop directly
Representation
点云 → 3D 高斯泼溅 → 可漫游场景Point cloud → 3D Gaussian splatting → walkable scene
Output
可控视频 · 三维场景 · 仿真数据Controlled video · 3D scenes · simulation data
Loop
Real-to-Sim → Sim-to-Real 数据飞轮Real-to-Sim → Sim-to-Real data flywheel
Benchmark
相机可控生成任务上，人类评审偏好领先专用模型Preferred over specialist models by human raters on camera-controlled generation
Floweriver World Model
03 — Product / 01
Floweriver World Model 空间智能世界模型The spatial-intelligence world model
四项能力统一于一个模型。给一段视频、一组相机位姿，或者一句描述，得到一段可以继续编辑的三维世界。 Four capabilities in one model. Feed it a video, a set of camera poses, or a sentence — get a 3D world you can keep editing.
- 相机位姿逐帧可控，支持关键帧与轨迹插值Frame-accurate camera control with keyframes and interpolation
- 单目视频重建为可漫游的 3D 高斯场景Monocular video into a roamable Gaussian scene
- 长时序生成保持几何一致，不漂移Geometrically consistent long-horizon generation
BayeSim · Simulation Engine
03 — Product / 02
BayeSim 仿真引擎Simulation engine
把真实世界转化为可规模化生产、可标注、可回放的仿真数据。真实一次采集，仿真无限次复现。 Turn the real world into scalable, labelled, replayable simulation data. Capture once, replay endlessly.
- Real-to-Sim：实景扫描直接生成可交互场景Real-to-Sim: real captures become interactive scenes
- 场景、光照、材质、相机批量参数化Batch-parameterised scenes, lighting, materials, cameras
- 为具身智能与自动驾驶输出训练集Training sets for embodied AI and autonomy
BAOLIMEI · AI-native Studio
03 — Product / 03
葆历莓 BAOLIMEI AI 原生创意工作室An AI-native creative studio
以原创故事与角色为核心，用世界模型做影像，孵化属于这个时代的 IP。 Original stories and characters at the core, filmed with world models — building the IP of this era.
- 原创角色与世界观资产库Original characters and world-building assets
- 世界模型驱动的短片与动画生产Short films and animation powered by world models
- 品牌联名与 IP 授权合作Brand collaborations and IP licensing
「生成的尽头不是像素，是世界。当模型理解了光、材质、镜头与时间，创作就从「猜一张图」变成「造一个空间」。」 “The end of generation isn’t pixels — it’s the world. Once a model grasps light, material, lens and time, creation stops guessing at an image and starts building a space.”
Location
广州 · 天河Guangzhou · Tianhe
研发与创作都在同一个地方发生：一个做世界模型，一个用世界模型做内容。 Research and creation happen in one place — one builds the world model, the other creates with it.
04 — About
一家把世界当作模型来研究的公司 A company that treats the world as a model
Company
广州流花科技有限公司Guangzhou Floweriver Technology Co., Ltd.
USCC
91440106MAEP601206统一社会信用代码Unified social credit code
Location
广东省广州市天河区Tianhe District, Guangzhou, Guangdong
Contact
199 4318 8815商务合作 · 内测申请Business & private beta
IP
注册商标 · 受理专利 · 原创品牌视觉资产Registered trademarks, pending patents, original brand assets
Mission
运用人工智能与数字世界技术，推动创意内容产业发展与人类文化表达进步。Advance the creative industry and human cultural expression through AI and digital-world technology.
Vision
成为连接 AI、原创 IP 与数字世界的创意科技公司。A creative-technology company connecting AI, original IP and digital worlds.
Principle
长期主义 · 第一性原理 · 商业飞轮Long-termism · First principles · Compounding flywheel
4Capabilities
3Product lines
2026Founded
1Mission
一起重建世界 Let's rebuild the world
内测申请、仿真数据合作与 IP 联名，欢迎随时联系。 Private beta, simulation data partnerships and IP collaborations — get in touch.
Private beta 已开放 · 广州天河 Private beta open · Tianhe, Guangzhou