M
MusicMaster4/umbra-25d
2.5D silhouette-and-diorama animation framework for AI agents: Python scene in, cinematic 1080p film out (Blender EEVEE on the GPU).
○checked 2026-11-20 — 命令里的包名与条目不一致,未作为可复制命令展示。
更新 2026-11-16
在 GitHub 查看 ↗链接已于 2026-11-17 核查 · HTTP 200 · 打开的是厂商自己的官网
装完之后怎么算装好了
- 装完该看到什么 — 命令正常结束时退出码为 0,终端里没有 "command not found" 或 "permission denied"。
- 一行自检 — 按该仓库 README 里给出的最小用法跑一次;能按 README 的描述输出就是通了。
- 答不出来怎么办 — 这个项目没有写排错说明,去仓库的 Issues 搜 install。
安装命令取自该仓库;以上判据只描述可观察到的结果,不替代项目自己的文档。
- ★ 数取值
- 0 ★ · MusicMaster4/umbra-25d · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-15 · 最近见到 2026-11-20 · 共见到 6 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 个来源 · repo
- 安装
- 安装命令见上方「安装与配置」。
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
MusicMaster4/umbra-25d;我们收录于 2026-11-15;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# umbra
以下摘录自该仓库的 README(片段)
2.5D silhouette-and-diorama animation, written in Python, built for AI agents.
Describe a scene in ~60 lines; umbra stages it in Blender, lights it, animates it and renders a
1080p film on your GPU. No Blender skills, no keyframing, no node graphs.
python
from umbra import
class Hello(Scene):
duration = 8
def build(s):
s.table(at=(0, 0)); s.chair(at=(0, -0.75)); s.lamp(at=(0, 0), kind="pendant")
a = s.person(at=(0, -0.75), color=WHITE).sit(pose="sit_table")
b = s.person(at=(3.5, 2.5), color=RED, build="woman", hair="long")
b.walk((0.9, 0.5), t=1, dur=3).pose("wave", t=4.2)
a.look_at(b.anchor("head"), t=2.5)
s.cam.set(pos=(4.5, -5.5, 3.2), look=(0.3, 0, 1), lens=45)
s.cam.orbit(25, t=0, dur=8)
s.label(b.anchor("above"), "Hello", t=4.4, out=7.4)
bash
python -m umbra render scenes/00_hello.py # -> out/00_hello.mp4
## Gallery
Ten 20-second films, each one a single Python file in scenes/. The first drafts of scenes 04–10
were written by fresh AI agents from AGENTS.md, in at most six contact-sheet iterations each;
their friction reports shaped the API you see here.
01 · Cabin fight — blueprint look, crowd, fight choreography, a dive through cloud
02 · Corcovado — dusk look, terrain, chase cam, explosions, falling wreck
03 · Dinner date — candle light, hand IK, shallow focus
04 · Crime scene — noir look, lantern beam, callout labels
05 · Rocket launch — countdown, fire and smoke riding a moving anchor
06 · Night chase — procedural city, headlights, route trails
07 · Warehouse — clay look, tilt-shift miniature
08 · Storm — water, rain, lightning, a sweeping beam
09 · Penalty — crowd, ball trajectory, broadcast-style graphics
10 · Orbit — no people: pure motion graphics in space
## Install
| need | version | notes |
| Blender | 4.5 LTS or newer (developed on 5.2) | found automatically, or set UMBRA_BLENDER=/path/to/blender |
| ffmpeg | any recent | on PATH (or pip install imageio-ffmpeg) |
| Python | 3.9+ | only runs the CLI; no packages to install |
| GPU | anything Blender's EEVEE runs on | NVIDIA / AMD / Intel / Apple; no CUDA or OptiX setup |
bash
git clone https://github.com/MusicMaster4/umbra-25d && cd umbra-25d
python -m umbra doctor # checks Blender + ffmpeg
python -m umbra sheet scenes/00_hello.py # a contact sheet in ~10 s
## For AI agents
Point your agent at the repo and say what film you want. AGENTS.md (also loaded through
CLAUDE.md) is a single page with the whole API, the working loop and the art-direction rules; an agent
does not need to read the source.
> Read AGENTS.md, then write scenes/lighthouse.py: a 20-second film of a fishing boat in a storm
> near a lighthouse. Iterate with python -m umbra sheet until it looks right, then render it.
The loop is built to be cheap in tokens:
| step | command | cost |
| see the whole film | python -m umbra sheet scene.py | one image (12 timestamped frames), ~10 s |
| check a moment | python -m umbra sheet scene.py -t 4.2 -t 4.6 -t 5.0 | one image |
| check detail | python -m umbra still scene.py -t 6.5 | one 1080p frame |
| check without rendering | python -m umbra lint scene.py | text only, no GPU, ~5 s |
| deliver | python -m umbra render scene.py | mp4; the ten films here took 3–20 min each (median 7) at 1080p on an RTX 3070 |
Describe a scene in ~60 lines; umbra stages it in Blender, lights it, animates it and renders a
1080p film on your GPU. No Blender skills, no keyframing, no node graphs.
python
from umbra import
class Hello(Scene):
duration = 8
def build(s):
s.table(at=(0, 0)); s.chair(at=(0, -0.75)); s.lamp(at=(0, 0), kind="pendant")
a = s.person(at=(0, -0.75), color=WHITE).sit(pose="sit_table")
b = s.person(at=(3.5, 2.5), color=RED, build="woman", hair="long")
b.walk((0.9, 0.5), t=1, dur=3).pose("wave", t=4.2)
a.look_at(b.anchor("head"), t=2.5)
s.cam.set(pos=(4.5, -5.5, 3.2), look=(0.3, 0, 1), lens=45)
s.cam.orbit(25, t=0, dur=8)
s.label(b.anchor("above"), "Hello", t=4.4, out=7.4)
bash
python -m umbra render scenes/00_hello.py # -> out/00_hello.mp4
## Gallery
Ten 20-second films, each one a single Python file in scenes/. The first drafts of scenes 04–10
were written by fresh AI agents from AGENTS.md, in at most six contact-sheet iterations each;
their friction reports shaped the API you see here.
01 · Cabin fight — blueprint look, crowd, fight choreography, a dive through cloud
02 · Corcovado — dusk look, terrain, chase cam, explosions, falling wreck
03 · Dinner date — candle light, hand IK, shallow focus
04 · Crime scene — noir look, lantern beam, callout labels
05 · Rocket launch — countdown, fire and smoke riding a moving anchor
06 · Night chase — procedural city, headlights, route trails
07 · Warehouse — clay look, tilt-shift miniature
08 · Storm — water, rain, lightning, a sweeping beam
09 · Penalty — crowd, ball trajectory, broadcast-style graphics
10 · Orbit — no people: pure motion graphics in space
## Install
| need | version | notes |
| Blender | 4.5 LTS or newer (developed on 5.2) | found automatically, or set UMBRA_BLENDER=/path/to/blender |
| ffmpeg | any recent | on PATH (or pip install imageio-ffmpeg) |
| Python | 3.9+ | only runs the CLI; no packages to install |
| GPU | anything Blender's EEVEE runs on | NVIDIA / AMD / Intel / Apple; no CUDA or OptiX setup |
bash
git clone https://github.com/MusicMaster4/umbra-25d && cd umbra-25d
python -m umbra doctor # checks Blender + ffmpeg
python -m umbra sheet scenes/00_hello.py # a contact sheet in ~10 s
## For AI agents
Point your agent at the repo and say what film you want. AGENTS.md (also loaded through
CLAUDE.md) is a single page with the whole API, the working loop and the art-direction rules; an agent
does not need to read the source.
> Read AGENTS.md, then write scenes/lighthouse.py: a 20-second film of a fishing boat in a storm
> near a lighthouse. Iterate with python -m umbra sheet until it looks right, then render it.
The loop is built to be cheap in tokens:
| step | command | cost |
| see the whole film | python -m umbra sheet scene.py | one image (12 timestamped frames), ~10 s |
| check a moment | python -m umbra sheet scene.py -t 4.2 -t 4.6 -t 5.0 | one image |
| check detail | python -m umbra still scene.py -t 6.5 | one 1080p frame |
| check without rendering | python -m umbra lint scene.py | text only, no GPU, ~5 s |
| deliver | python -m umbra render scene.py | mp4; the ten films here took 3–20 min each (median 7) at 1080p on an RTX 3070 |
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/musicmaster4-umbra-25d/)
- 来源
- github · https://github.com/MusicMaster4/umbra-25d
- 按领域浏览
- 开源项目
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。