VideoCoCo 是一个物理忠实的视频生成管线,它使用代码作为思维链,先模拟并渲染一个携带正确物理特性的中性代理视频,再将其重写为逼真的风格。

Stars

4

7 天增长

暂无数据

Fork 数

2

开放 Issue

1

开源协议

暂无数据

最近更新

2026-07-29

AI 仓库情报摘要
FR-AI / ANALYSIS

为什么值得关注

它显式地在生成像素之前进行物理推理,通过代码代理生成仿真并验证因果关系,实现了物理正确性与视觉风格的解耦,并提供了可验证的草稿-编辑流程。

适合谁使用

  • 视频生成与物理仿真领域的研究人员
  • 关注思维链在视觉任务中应用的 AI 从业者
  • 需要物理合理视频编辑的内容创作者
  • 探索基于仿真管线的计算机视觉工程师

典型使用场景

  • 生成具有准确浮力、融化或弹性等物理效果的短视频
  • 编辑现有视频,为其赋予物理正确的运动同时保留风格
  • 通过代码仿真快速原型物理场景,再通过重写实现逼真效果
  • 作为评估视频生成模型物理忠实度的基准

项目优势

  • 通过代码代理和沙箱渲染显式进行物理推理,再生成像素
  • 验证步骤确保仿真与计划之间的因果一致性
  • 推理代码、玩具数据集和调优权重均开源并提供于 Hugging Face Hub
  • 模块化管线包含五个代理技能,每个专注于清晰子任务

使用前须知

  • 玩具数据集仅包含 8 个手工检查的案例,不足以进行训练规模评估
  • 依赖外部 OmniWeaving 模型及腾讯混元社区许可协议
  • 调优权重仍在上传中(按 README 说明),可能暂时无法使用

README 快速开始

🎬 VideoCoCo: Code as CoT for Physics-Faithful Video Generation

Official repository for VideoCoCo, a physics-faithful video generation pipeline that uses code as a chain-of-thought to draft physics before committing to pixels.

[🤗 Weights] [💻 Code]

💥 News

  • [2026.07.29] We release the Agent Skills, a toy dataset, and the inference code. Tuned weights are uploading to the 🤗 Hub.

🪄 Draft Before Generation

We propose VideoCoCo, an interleaved reasoning paradigm that carries a physical prior through an explicit visual draft before committing to pixels.

Our method 🎨 first has a code agent write simulation code and render it in a sandbox as a neutral white/clay proxy video that carries the correct motion, causality, and physics — meaning is expressed by shape, transparency, deformation, and coverage, never by color.

Then we 🔎 verify the proxy against the physical plan (a caused state must stay hidden until its causing transition), and 🖼️ restyle the proxy into a photorealistic video driven by a per-case edit instruction.

📦 What's in this repo

  • skill/ — the five Agent Skills forming the pipeline: physical-state-plannerphysical-video-blender-implementerblender-mcp-videoseedance-edit-promptseedance-distill.
  • data/toy_cases/ — 8 hand-checked video-to-video (v2v) triplets.
  • inference/ — batch inference scripts + a patch against upstream OmniWeaving.
  • 🤗 mickyhimself/VideoCoCo — the tuned transformer.

🎬 Toy dataset

data/toy_cases/ — 8 v2v triplets, one directory per case:

data/toy_cases/
├── manifest.jsonl                 # one JSON line per case (index)
├── 0000_buoyancy/
│   ├── video.mp4                  # source: neutral white/clay physics proxy
│   ├── seedance.mp4               # target: photoreal restyle
│   └── edit_prompt.txt            # instruction used to restyle proxy -> photoreal
└── ...

Each manifest.jsonl line:

{"case_id": "0000_buoyancy", "source": "0000_buoyancy/video.mp4", "target": "0000_buoyancy/seedance.mp4", "instruction": "...", "category": "buoyancy"}
  • source (video.mp4) — a grayscale/white-material rend

项目描述

Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

相关仓库与替代方案

根据分类、Topic 和编程语言匹配的相似项目。

gnipbao
精选
gnipbao GitHub avatar

story-to-handdrawn-video

A Remotion-based tool that converts Chinese story text or ordered hand-drawn images into vertical hand-drawn diary-comic animation with handwritten captions, left-to-right reveals, optional page-curl transitions, and silent H.264 output for post-production dubbing.

AI 与机器学习AI 智能体
683
MoonshotAI
精选
MoonshotAI GitHub avatar

Kimi-K3

Kimi K3 is an open-weight, 2.8T-parameter native multimodal agentic model with a 1M-token context window, designed for frontier coding, knowledge work, and reasoning tasks.

AI 与机器学习AI 智能体
3,348
yetone
精选
yetone GitHub avatar

kill-ai-slop

Kill AI Slop is a multilingual field guide and agent skill that catalogues 33 common AI-generated design tells with interactive before-after demos, and provides a tool to scan and fix web projects.

AI 与机器学习AI 智能体
805