文字没有被参考图替代
参考图锁住可见事实，文字负责说明这一镜为什么存在、什么能变、什么不能变。
Reference
Lab
关于 AI 视频、短剧生产、提示词系统和视觉方法的实验记录。
Featured Note
文字没有被参考图替代
参考图锁住可见事实，文字负责说明这一镜为什么存在、什么能变、什么不能变。
场景记忆不是氛围
氛围词可以保留调性，但不能替模型记住入口、窗、桌子、道具和光线方向。
Continuity
Seedance 不是分镜生成器
模型可以生成镜头，但不能替上游做取舍；提示词只应该交付当前镜头需要执行的决定。
Seedance
Structural Pass Is Not Production Pass
A field note on semantic QA, production readiness, and render-feedback routing for AI video prompt systems.
Production Notes
Seedance 2.0: Multi-shot Becomes Director Control
A note on Seedance as a move from single-clip generation toward multimodal shot-flow control.
AI
Kling 3.0: Reference, Audio, and Storyboard Control
A note on Kling 3.0 as a shift from isolated clips toward reference-driven sequence control.
AI
Hailuo: Performance, Motion, and Cost
A note on why practical video models compete on repeatable performance, not only peak demos.
Tools
HunyuanVideo and Wan: Open Models Become Infrastructure
A note on open video stacks, first-last-frame control, and why inspectability matters.
Research
Veo 2: Cinematography Becomes Promptable
A note on why Veo 2's real shift was camera behavior, physics, and film grammar.
AI
Luma Dream Machine: Video as a Creative Instrument
A note on keyframes, camera control, extension, and why iteration surface changes model value.
Tools
Character Placement and Continuity Tests
A practical note on identity, placement, and relationship anchors across shots.
Production Notes
Prompt Structure and Scene Rhythm
How character assets become shot prompts without overloading the model.
Workflow
Building Reference Sets for Characters
Reference curation as a system for identity, silhouette, costume, posture, and emotional range.
Pipeline
Spatial Anchors: Terrain, Light, and Time
Using place, direction, light, and repeated objects as scaffolding for generated scenes.
Visual Systems
From Script to Shot Flow
Turning narrative beats into ordered visual states, not just a list of pretty shots.
Process
Camera Language in AI Video
Framing, lens behavior, and movement as continuity tools, not decorative cinema words.
Cinematography
Sora: The World-Simulator Frame
A note on why Sora changed the question from clip generation to visual state simulation.
AI
Lumiere: Generating the Whole Motion at Once
A note on whole-clip generation as an answer to global temporal consistency.
Research
Stable Video Diffusion: The Open-Weights Pivot
A note on why open weights changed video generation from a demo into material for inspection.
Research
Runway Gen-2: The Browser Studio Moment
A note on why Gen-2 mattered as a creative interface, not only as a model.
Tools
Imagen Video and Phenaki: Fidelity or Time
A note on the early split between sharper video and longer narrative continuity.
Research
Make-A-Video: Learning Motion Without Paired Captions
A note on the data problem behind early text-to-video generation.
Research
CogVideo: Inheriting Image Models for Motion
A note on the early strategy of borrowing image-model semantics before learning motion.
Research
Video Diffusion Models: Motion Enters Diffusion
A note on why video generation first had to become a problem of time, not image quality.
Research