mediakit-cli 0.1.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- mediakit_cli-0.1.2/LICENSE +21 -0
- mediakit_cli-0.1.2/PKG-INFO +221 -0
- mediakit_cli-0.1.2/README.md +162 -0
- mediakit_cli-0.1.2/mediakit_cli/__init__.py +5 -0
- mediakit_cli-0.1.2/mediakit_cli/__main__.py +4 -0
- mediakit_cli-0.1.2/mediakit_cli/auto.py +217 -0
- mediakit_cli-0.1.2/mediakit_cli/cli.py +665 -0
- mediakit_cli-0.1.2/mediakit_cli/config.py +68 -0
- mediakit_cli-0.1.2/mediakit_cli/dub.py +174 -0
- mediakit_cli-0.1.2/mediakit_cli/edit.py +280 -0
- mediakit_cli-0.1.2/mediakit_cli/index.py +101 -0
- mediakit_cli-0.1.2/mediakit_cli/lyrics.py +92 -0
- mediakit_cli-0.1.2/mediakit_cli/mcp_server.py +162 -0
- mediakit_cli-0.1.2/mediakit_cli/music.py +226 -0
- mediakit_cli-0.1.2/mediakit_cli/ocr.py +129 -0
- mediakit_cli-0.1.2/mediakit_cli/polish.py +85 -0
- mediakit_cli-0.1.2/mediakit_cli/probe.py +89 -0
- mediakit_cli-0.1.2/mediakit_cli/shotcard.py +133 -0
- mediakit_cli-0.1.2/mediakit_cli/shots.py +109 -0
- mediakit_cli-0.1.2/mediakit_cli/subs.py +349 -0
- mediakit_cli-0.1.2/mediakit_cli/util.py +111 -0
- mediakit_cli-0.1.2/mediakit_cli/vlm.py +88 -0
- mediakit_cli-0.1.2/mediakit_cli/watch.py +119 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/PKG-INFO +221 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/SOURCES.txt +33 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/dependency_links.txt +1 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/entry_points.txt +3 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/requires.txt +57 -0
- mediakit_cli-0.1.2/mediakit_cli.egg-info/top_level.txt +1 -0
- mediakit_cli-0.1.2/pyproject.toml +75 -0
- mediakit_cli-0.1.2/setup.cfg +4 -0
- mediakit_cli-0.1.2/tests/test_auto.py +32 -0
- mediakit_cli-0.1.2/tests/test_config.py +29 -0
- mediakit_cli-0.1.2/tests/test_edit_math.py +28 -0
- mediakit_cli-0.1.2/tests/test_subs.py +44 -0
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 mediakit contributors
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,221 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: mediakit-cli
|
|
3
|
+
Version: 0.1.2
|
|
4
|
+
Summary: Agent-friendly CLI for watching videos, analyzing music, subtitling, dubbing and beat-synced editing.
|
|
5
|
+
Author: mediakit contributors
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Project-URL: Homepage, https://github.com/your-name/mediakit
|
|
8
|
+
Keywords: video,subtitles,ffmpeg,whisper,agent,mcp
|
|
9
|
+
Classifier: Programming Language :: Python :: 3
|
|
10
|
+
Classifier: Operating System :: OS Independent
|
|
11
|
+
Classifier: Topic :: Multimedia :: Video
|
|
12
|
+
Classifier: Topic :: Multimedia :: Sound/Audio
|
|
13
|
+
Requires-Python: >=3.10
|
|
14
|
+
Description-Content-Type: text/markdown
|
|
15
|
+
License-File: LICENSE
|
|
16
|
+
Requires-Dist: numpy>=1.24
|
|
17
|
+
Requires-Dist: pillow>=10
|
|
18
|
+
Requires-Dist: pysubs2>=1.6
|
|
19
|
+
Provides-Extra: asr
|
|
20
|
+
Requires-Dist: faster-whisper>=1.0; extra == "asr"
|
|
21
|
+
Provides-Extra: vision
|
|
22
|
+
Requires-Dist: opencv-python>=4.8; extra == "vision"
|
|
23
|
+
Requires-Dist: scenedetect>=0.6; extra == "vision"
|
|
24
|
+
Requires-Dist: rapidocr-onnxruntime>=1.3; extra == "vision"
|
|
25
|
+
Provides-Extra: timeline
|
|
26
|
+
Requires-Dist: opentimelineio>=0.16; extra == "timeline"
|
|
27
|
+
Provides-Extra: dl
|
|
28
|
+
Requires-Dist: yt-dlp>=2024.1.1; extra == "dl"
|
|
29
|
+
Provides-Extra: music
|
|
30
|
+
Requires-Dist: librosa>=0.10; extra == "music"
|
|
31
|
+
Requires-Dist: scipy>=1.10; extra == "music"
|
|
32
|
+
Requires-Dist: soundfile>=0.12; extra == "music"
|
|
33
|
+
Provides-Extra: separate
|
|
34
|
+
Requires-Dist: audio-separator>=0.20; extra == "separate"
|
|
35
|
+
Provides-Extra: search
|
|
36
|
+
Requires-Dist: open-clip-torch>=2.20; extra == "search"
|
|
37
|
+
Provides-Extra: mcp
|
|
38
|
+
Requires-Dist: mcp>=1.0; extra == "mcp"
|
|
39
|
+
Provides-Extra: windows
|
|
40
|
+
Requires-Dist: pywin32>=306; sys_platform == "win32" and extra == "windows"
|
|
41
|
+
Provides-Extra: all
|
|
42
|
+
Requires-Dist: faster-whisper>=1.0; extra == "all"
|
|
43
|
+
Requires-Dist: opencv-python>=4.8; extra == "all"
|
|
44
|
+
Requires-Dist: scenedetect>=0.6; extra == "all"
|
|
45
|
+
Requires-Dist: rapidocr-onnxruntime>=1.3; extra == "all"
|
|
46
|
+
Requires-Dist: opentimelineio>=0.16; extra == "all"
|
|
47
|
+
Requires-Dist: yt-dlp>=2024.1.1; extra == "all"
|
|
48
|
+
Requires-Dist: librosa>=0.10; extra == "all"
|
|
49
|
+
Requires-Dist: scipy>=1.10; extra == "all"
|
|
50
|
+
Requires-Dist: soundfile>=0.12; extra == "all"
|
|
51
|
+
Requires-Dist: audio-separator>=0.20; extra == "all"
|
|
52
|
+
Requires-Dist: open-clip-torch>=2.20; extra == "all"
|
|
53
|
+
Requires-Dist: mcp>=1.0; extra == "all"
|
|
54
|
+
Requires-Dist: pywin32>=306; sys_platform == "win32" and extra == "all"
|
|
55
|
+
Provides-Extra: dev
|
|
56
|
+
Requires-Dist: pytest>=7.4; extra == "dev"
|
|
57
|
+
Requires-Dist: ruff>=0.4; extra == "dev"
|
|
58
|
+
Dynamic: license-file
|
|
59
|
+
|
|
60
|
+
# mediakit
|
|
61
|
+
|
|
62
|
+
**One CLI, 17 commands, JSON out — for agents that need to watch, listen, subtitle and cut.**
|
|
63
|
+
|
|
64
|
+
`mediakit` is a thin orchestration layer over mature media tools. It does not
|
|
65
|
+
reimplement decoding, ASR or scene detection; it calls them and returns
|
|
66
|
+
structured results an agent can reason about, plus an MCP server so the whole
|
|
67
|
+
toolkit can be attached to any MCP client.
|
|
68
|
+
|
|
69
|
+
```
|
|
70
|
+
$ mk watch clip.mp4 --out work/clip --lang zh --ocr
|
|
71
|
+
语言 zh · 4 镜头 · 12 关键帧 · 转写 8 段 · 画面文字 2 条
|
|
72
|
+
$ mk search work/clip/index "穿黑裙子的女孩在草地上" --top 3
|
|
73
|
+
42.50s score 0.311 ...
|
|
74
|
+
```
|
|
75
|
+
|
|
76
|
+
[中文说明](README.zh-CN.md)
|
|
77
|
+
|
|
78
|
+
## Why another wrapper?
|
|
79
|
+
|
|
80
|
+
Most tools are excellent at one step and then stop: PySceneDetect prints a
|
|
81
|
+
table, Whisper emits one long line, CLIP libraries return tensors. An agent
|
|
82
|
+
needs the opposite: short, composable commands with stable JSON, and no GUI.
|
|
83
|
+
|
|
84
|
+
| Upstream | What it does | What `mk` adds |
|
|
85
|
+
|---|---|---|
|
|
86
|
+
| PySceneDetect | shot boundaries | `contact.jpg`, `shots.json`, OTIO export |
|
|
87
|
+
| faster-whisper | segments + words | pause-based line breaking, width limits, 繁→简, karaoke, QA |
|
|
88
|
+
| ffmpeg | encode/cut/burn | one plan file, xfade + acrossfade, subtitle rescaling |
|
|
89
|
+
| RapidOCR | text boxes | frame sampling, de-dup, SRT output |
|
|
90
|
+
| CLIP | embeddings | index/search commands with thumbnails |
|
|
91
|
+
| librosa | tempo/beats | tempogram tempo prior, sections, chords, melody |
|
|
92
|
+
| VideoCaptioner | LLM subtitle polish | one command, local Ollama, bilingual output |
|
|
93
|
+
|
|
94
|
+
## Install
|
|
95
|
+
|
|
96
|
+
```bash
|
|
97
|
+
pip install "mediakit-cli[all]" # everything
|
|
98
|
+
pip install mediakit-cli # core only: probe, edit, auto, plan logic
|
|
99
|
+
```
|
|
100
|
+
|
|
101
|
+
`ffmpeg` must be on `PATH` (or set `MEDIAKIT_FFMPEG`). Optional groups:
|
|
102
|
+
|
|
103
|
+
The distribution is named `mediakit-cli`; the Python import package is
|
|
104
|
+
`mediakit_cli` (to avoid a collision with the unrelated PyPI package
|
|
105
|
+
`mediakit`). The CLI command is `mk` either way.
|
|
106
|
+
|
|
107
|
+
| Extra | Enables | Install |
|
|
108
|
+
|---|---|---|
|
|
109
|
+
| `asr` | `subs`, `watch`, `lyrics` | `faster-whisper` |
|
|
110
|
+
| `vision` | `shots`, `shotcard`, `ocr` | OpenCV, PySceneDetect, RapidOCR |
|
|
111
|
+
| `music` | `music` | librosa |
|
|
112
|
+
| `search` | `index`, `search` | open_clip + ViT-B-32 weights |
|
|
113
|
+
| `separate` | `lyrics --separate` | audio-separator |
|
|
114
|
+
| `mcp` | `mediakit-mcp` | MCP SDK |
|
|
115
|
+
| `windows` | `dub`, `voices` | pywin32 + SAPI |
|
|
116
|
+
| `timeline` | `shots --otio` | OpenTimelineIO |
|
|
117
|
+
|
|
118
|
+
Environment overrides: `MEDIAKIT_FFMPEG`, `MEDIAKIT_AUDIO_SEPARATOR`,
|
|
119
|
+
`MEDIAKIT_VIDEOCAPTIONER`, `MEDIAKIT_OLLAMA_BASE`, `MEDIAKIT_VLM_MODEL`.
|
|
120
|
+
|
|
121
|
+
## Commands
|
|
122
|
+
|
|
123
|
+
```bash
|
|
124
|
+
# look
|
|
125
|
+
mk probe a.mp4
|
|
126
|
+
mk watch a.mp4 --out work/a --lang zh --ocr
|
|
127
|
+
mk frames a.mp4 --n 12 --out work/a/frames
|
|
128
|
+
mk ocr a.mp4 --out work/a/ocr --every 2
|
|
129
|
+
mk index a.mp4 --out work/a/index
|
|
130
|
+
mk search work/a/index "a girl running on grass"
|
|
131
|
+
|
|
132
|
+
# listen
|
|
133
|
+
mk music song.mp3 --chords --melody --engine beat-this
|
|
134
|
+
mk lyrics song.mp3 --out work/lyrics --lang ja
|
|
135
|
+
|
|
136
|
+
# dissect
|
|
137
|
+
mk shots a.mp4 --out work/a --otio
|
|
138
|
+
mk shotcard a.mp4 --out work/a/card --vlm
|
|
139
|
+
|
|
140
|
+
# subtitle
|
|
141
|
+
mk subs a.mp4 --lang zh --out work/a --size 1080x1920 --max-chars 12 --karaoke --qa
|
|
142
|
+
mk subqa work/a/subs.srt
|
|
143
|
+
mk polish work/a/subs.srt --out work/a/polished.srt
|
|
144
|
+
mk voices
|
|
145
|
+
mk dub a.mp4 --srt work/a/subs.srt --out work/dub.mp4 --voice Huihui --mix-original
|
|
146
|
+
|
|
147
|
+
# cut
|
|
148
|
+
mk edit work/plan.json --out work/final.mp4
|
|
149
|
+
mk auto --music bgm.mp3 --src footage --out work/mv.mp4 --transition fade
|
|
150
|
+
```
|
|
151
|
+
|
|
152
|
+
Every command accepts `--json`. Without it you get a short human summary.
|
|
153
|
+
|
|
154
|
+
### `mk auto` — beat-synced montage
|
|
155
|
+
|
|
156
|
+
Analyzes the track, decides how many beats per cut per section (quiet sections
|
|
157
|
+
get longer cuts, loud ones shorter), picks shots from a folder, snaps cuts to
|
|
158
|
+
beats and renders `xfade` transitions with matching `acrossfade` audio.
|
|
159
|
+
|
|
160
|
+
```bash
|
|
161
|
+
mk auto --music bgm.mp3 --src footage --out mv.mp4 \
|
|
162
|
+
--size 1080x1920 --fit crop --transition fade --seed 7
|
|
163
|
+
mk auto --music bgm.mp3 --src footage --out mv.mp4 --plan-only # plan only
|
|
164
|
+
```
|
|
165
|
+
|
|
166
|
+
## MCP server
|
|
167
|
+
|
|
168
|
+
```json
|
|
169
|
+
{
|
|
170
|
+
"mcpServers": {
|
|
171
|
+
"mediakit": { "command": "mediakit-mcp" }
|
|
172
|
+
}
|
|
173
|
+
}
|
|
174
|
+
```
|
|
175
|
+
|
|
176
|
+
12 tools: `probe`, `shots`, `music`, `subs`, `ocr`, `shotcard`, `index`,
|
|
177
|
+
`search`, `lyrics`, `dub`, `edit`, `auto`.
|
|
178
|
+
|
|
179
|
+
## Measured, not assumed
|
|
180
|
+
|
|
181
|
+
Numbers below were produced on a local machine (CPU) with the benchmark
|
|
182
|
+
scripts in `tools/`. They are small-sample results, not leaderboard claims.
|
|
183
|
+
|
|
184
|
+
| Area | Result |
|
|
185
|
+
|---|---|
|
|
186
|
+
| Shot detection | 3 hard cuts detected at exactly 4.00s / 8.00s |
|
|
187
|
+
| Subtitles | 16.5s Chinese speech → 9 lines, traditional→simplified, width capped |
|
|
188
|
+
| Karaoke | `\kf` per word; at 0.2s 65% of subtitle pixels still unsung color |
|
|
189
|
+
| OCR | read burned subtitles back as 「大家晚上好。」and merged into SRT |
|
|
190
|
+
| Retrieval (synthetic, 8 segments) | recall@1 75% descriptive / 88% short, recall@3 100% |
|
|
191
|
+
| Retrieval (real footage, 8 clips) | recall@1 75%, recall@3 100% |
|
|
192
|
+
| Lyric CER vs reference subs | mix 89.7% → vocals 60.3% (small) → **54.4% (medium)** |
|
|
193
|
+
| Tempo correction | librosa default 103.36 BPM was a 4:3 artifact; tempogram prior → 80.75 BPM |
|
|
194
|
+
| Montage | 7 segments from a 20s track; 18.50s with 6×0.25s fades |
|
|
195
|
+
| Audio crossfade | 3×2s clips + 0.25s fade + keep_audio → 5.50s, audio kept |
|
|
196
|
+
| Dubbing | 9 lines → 18.60s output, full length, 2ch/48kHz |
|
|
197
|
+
|
|
198
|
+
## Known limits
|
|
199
|
+
|
|
200
|
+
- Retrieval benchmarks are small (8 segments/clips) and CLIP is
|
|
201
|
+
phrasing-sensitive; low-information frames (pure white/black) fail.
|
|
202
|
+
- Singing transcription is still hard: 54% CER is the good case.
|
|
203
|
+
- `dub` uses Windows SAPI — robotic, and long lines are time-compressed at most 2×.
|
|
204
|
+
- VLM shot descriptions were validated on a handful of frames.
|
|
205
|
+
- `beat_this` and librosa often disagree by a few BPM; the package now anchors
|
|
206
|
+
tempo with a tempogram prior but does not prove which tracker is musically right.
|
|
207
|
+
- Transition handling assumes hard cuts when detecting shots.
|
|
208
|
+
|
|
209
|
+
## Layout
|
|
210
|
+
|
|
211
|
+
```
|
|
212
|
+
mediakit_cli/ package (17 modules + MCP server)
|
|
213
|
+
tools/ benchmarks: retrieval, lyrics CER, tempo
|
|
214
|
+
docs/ architecture and data contracts
|
|
215
|
+
examples/ sample edit plan
|
|
216
|
+
tests/ unit tests for pure logic
|
|
217
|
+
```
|
|
218
|
+
|
|
219
|
+
## License
|
|
220
|
+
|
|
221
|
+
MIT
|
|
@@ -0,0 +1,162 @@
|
|
|
1
|
+
# mediakit
|
|
2
|
+
|
|
3
|
+
**One CLI, 17 commands, JSON out — for agents that need to watch, listen, subtitle and cut.**
|
|
4
|
+
|
|
5
|
+
`mediakit` is a thin orchestration layer over mature media tools. It does not
|
|
6
|
+
reimplement decoding, ASR or scene detection; it calls them and returns
|
|
7
|
+
structured results an agent can reason about, plus an MCP server so the whole
|
|
8
|
+
toolkit can be attached to any MCP client.
|
|
9
|
+
|
|
10
|
+
```
|
|
11
|
+
$ mk watch clip.mp4 --out work/clip --lang zh --ocr
|
|
12
|
+
语言 zh · 4 镜头 · 12 关键帧 · 转写 8 段 · 画面文字 2 条
|
|
13
|
+
$ mk search work/clip/index "穿黑裙子的女孩在草地上" --top 3
|
|
14
|
+
42.50s score 0.311 ...
|
|
15
|
+
```
|
|
16
|
+
|
|
17
|
+
[中文说明](README.zh-CN.md)
|
|
18
|
+
|
|
19
|
+
## Why another wrapper?
|
|
20
|
+
|
|
21
|
+
Most tools are excellent at one step and then stop: PySceneDetect prints a
|
|
22
|
+
table, Whisper emits one long line, CLIP libraries return tensors. An agent
|
|
23
|
+
needs the opposite: short, composable commands with stable JSON, and no GUI.
|
|
24
|
+
|
|
25
|
+
| Upstream | What it does | What `mk` adds |
|
|
26
|
+
|---|---|---|
|
|
27
|
+
| PySceneDetect | shot boundaries | `contact.jpg`, `shots.json`, OTIO export |
|
|
28
|
+
| faster-whisper | segments + words | pause-based line breaking, width limits, 繁→简, karaoke, QA |
|
|
29
|
+
| ffmpeg | encode/cut/burn | one plan file, xfade + acrossfade, subtitle rescaling |
|
|
30
|
+
| RapidOCR | text boxes | frame sampling, de-dup, SRT output |
|
|
31
|
+
| CLIP | embeddings | index/search commands with thumbnails |
|
|
32
|
+
| librosa | tempo/beats | tempogram tempo prior, sections, chords, melody |
|
|
33
|
+
| VideoCaptioner | LLM subtitle polish | one command, local Ollama, bilingual output |
|
|
34
|
+
|
|
35
|
+
## Install
|
|
36
|
+
|
|
37
|
+
```bash
|
|
38
|
+
pip install "mediakit-cli[all]" # everything
|
|
39
|
+
pip install mediakit-cli # core only: probe, edit, auto, plan logic
|
|
40
|
+
```
|
|
41
|
+
|
|
42
|
+
`ffmpeg` must be on `PATH` (or set `MEDIAKIT_FFMPEG`). Optional groups:
|
|
43
|
+
|
|
44
|
+
The distribution is named `mediakit-cli`; the Python import package is
|
|
45
|
+
`mediakit_cli` (to avoid a collision with the unrelated PyPI package
|
|
46
|
+
`mediakit`). The CLI command is `mk` either way.
|
|
47
|
+
|
|
48
|
+
| Extra | Enables | Install |
|
|
49
|
+
|---|---|---|
|
|
50
|
+
| `asr` | `subs`, `watch`, `lyrics` | `faster-whisper` |
|
|
51
|
+
| `vision` | `shots`, `shotcard`, `ocr` | OpenCV, PySceneDetect, RapidOCR |
|
|
52
|
+
| `music` | `music` | librosa |
|
|
53
|
+
| `search` | `index`, `search` | open_clip + ViT-B-32 weights |
|
|
54
|
+
| `separate` | `lyrics --separate` | audio-separator |
|
|
55
|
+
| `mcp` | `mediakit-mcp` | MCP SDK |
|
|
56
|
+
| `windows` | `dub`, `voices` | pywin32 + SAPI |
|
|
57
|
+
| `timeline` | `shots --otio` | OpenTimelineIO |
|
|
58
|
+
|
|
59
|
+
Environment overrides: `MEDIAKIT_FFMPEG`, `MEDIAKIT_AUDIO_SEPARATOR`,
|
|
60
|
+
`MEDIAKIT_VIDEOCAPTIONER`, `MEDIAKIT_OLLAMA_BASE`, `MEDIAKIT_VLM_MODEL`.
|
|
61
|
+
|
|
62
|
+
## Commands
|
|
63
|
+
|
|
64
|
+
```bash
|
|
65
|
+
# look
|
|
66
|
+
mk probe a.mp4
|
|
67
|
+
mk watch a.mp4 --out work/a --lang zh --ocr
|
|
68
|
+
mk frames a.mp4 --n 12 --out work/a/frames
|
|
69
|
+
mk ocr a.mp4 --out work/a/ocr --every 2
|
|
70
|
+
mk index a.mp4 --out work/a/index
|
|
71
|
+
mk search work/a/index "a girl running on grass"
|
|
72
|
+
|
|
73
|
+
# listen
|
|
74
|
+
mk music song.mp3 --chords --melody --engine beat-this
|
|
75
|
+
mk lyrics song.mp3 --out work/lyrics --lang ja
|
|
76
|
+
|
|
77
|
+
# dissect
|
|
78
|
+
mk shots a.mp4 --out work/a --otio
|
|
79
|
+
mk shotcard a.mp4 --out work/a/card --vlm
|
|
80
|
+
|
|
81
|
+
# subtitle
|
|
82
|
+
mk subs a.mp4 --lang zh --out work/a --size 1080x1920 --max-chars 12 --karaoke --qa
|
|
83
|
+
mk subqa work/a/subs.srt
|
|
84
|
+
mk polish work/a/subs.srt --out work/a/polished.srt
|
|
85
|
+
mk voices
|
|
86
|
+
mk dub a.mp4 --srt work/a/subs.srt --out work/dub.mp4 --voice Huihui --mix-original
|
|
87
|
+
|
|
88
|
+
# cut
|
|
89
|
+
mk edit work/plan.json --out work/final.mp4
|
|
90
|
+
mk auto --music bgm.mp3 --src footage --out work/mv.mp4 --transition fade
|
|
91
|
+
```
|
|
92
|
+
|
|
93
|
+
Every command accepts `--json`. Without it you get a short human summary.
|
|
94
|
+
|
|
95
|
+
### `mk auto` — beat-synced montage
|
|
96
|
+
|
|
97
|
+
Analyzes the track, decides how many beats per cut per section (quiet sections
|
|
98
|
+
get longer cuts, loud ones shorter), picks shots from a folder, snaps cuts to
|
|
99
|
+
beats and renders `xfade` transitions with matching `acrossfade` audio.
|
|
100
|
+
|
|
101
|
+
```bash
|
|
102
|
+
mk auto --music bgm.mp3 --src footage --out mv.mp4 \
|
|
103
|
+
--size 1080x1920 --fit crop --transition fade --seed 7
|
|
104
|
+
mk auto --music bgm.mp3 --src footage --out mv.mp4 --plan-only # plan only
|
|
105
|
+
```
|
|
106
|
+
|
|
107
|
+
## MCP server
|
|
108
|
+
|
|
109
|
+
```json
|
|
110
|
+
{
|
|
111
|
+
"mcpServers": {
|
|
112
|
+
"mediakit": { "command": "mediakit-mcp" }
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
```
|
|
116
|
+
|
|
117
|
+
12 tools: `probe`, `shots`, `music`, `subs`, `ocr`, `shotcard`, `index`,
|
|
118
|
+
`search`, `lyrics`, `dub`, `edit`, `auto`.
|
|
119
|
+
|
|
120
|
+
## Measured, not assumed
|
|
121
|
+
|
|
122
|
+
Numbers below were produced on a local machine (CPU) with the benchmark
|
|
123
|
+
scripts in `tools/`. They are small-sample results, not leaderboard claims.
|
|
124
|
+
|
|
125
|
+
| Area | Result |
|
|
126
|
+
|---|---|
|
|
127
|
+
| Shot detection | 3 hard cuts detected at exactly 4.00s / 8.00s |
|
|
128
|
+
| Subtitles | 16.5s Chinese speech → 9 lines, traditional→simplified, width capped |
|
|
129
|
+
| Karaoke | `\kf` per word; at 0.2s 65% of subtitle pixels still unsung color |
|
|
130
|
+
| OCR | read burned subtitles back as 「大家晚上好。」and merged into SRT |
|
|
131
|
+
| Retrieval (synthetic, 8 segments) | recall@1 75% descriptive / 88% short, recall@3 100% |
|
|
132
|
+
| Retrieval (real footage, 8 clips) | recall@1 75%, recall@3 100% |
|
|
133
|
+
| Lyric CER vs reference subs | mix 89.7% → vocals 60.3% (small) → **54.4% (medium)** |
|
|
134
|
+
| Tempo correction | librosa default 103.36 BPM was a 4:3 artifact; tempogram prior → 80.75 BPM |
|
|
135
|
+
| Montage | 7 segments from a 20s track; 18.50s with 6×0.25s fades |
|
|
136
|
+
| Audio crossfade | 3×2s clips + 0.25s fade + keep_audio → 5.50s, audio kept |
|
|
137
|
+
| Dubbing | 9 lines → 18.60s output, full length, 2ch/48kHz |
|
|
138
|
+
|
|
139
|
+
## Known limits
|
|
140
|
+
|
|
141
|
+
- Retrieval benchmarks are small (8 segments/clips) and CLIP is
|
|
142
|
+
phrasing-sensitive; low-information frames (pure white/black) fail.
|
|
143
|
+
- Singing transcription is still hard: 54% CER is the good case.
|
|
144
|
+
- `dub` uses Windows SAPI — robotic, and long lines are time-compressed at most 2×.
|
|
145
|
+
- VLM shot descriptions were validated on a handful of frames.
|
|
146
|
+
- `beat_this` and librosa often disagree by a few BPM; the package now anchors
|
|
147
|
+
tempo with a tempogram prior but does not prove which tracker is musically right.
|
|
148
|
+
- Transition handling assumes hard cuts when detecting shots.
|
|
149
|
+
|
|
150
|
+
## Layout
|
|
151
|
+
|
|
152
|
+
```
|
|
153
|
+
mediakit_cli/ package (17 modules + MCP server)
|
|
154
|
+
tools/ benchmarks: retrieval, lyrics CER, tempo
|
|
155
|
+
docs/ architecture and data contracts
|
|
156
|
+
examples/ sample edit plan
|
|
157
|
+
tests/ unit tests for pure logic
|
|
158
|
+
```
|
|
159
|
+
|
|
160
|
+
## License
|
|
161
|
+
|
|
162
|
+
MIT
|
|
@@ -0,0 +1,217 @@
|
|
|
1
|
+
from __future__ import annotations
|
|
2
|
+
|
|
3
|
+
import json
|
|
4
|
+
import random
|
|
5
|
+
from pathlib import Path
|
|
6
|
+
|
|
7
|
+
from .edit import render
|
|
8
|
+
from .music import analyze
|
|
9
|
+
from .util import ensure_dir, media_path
|
|
10
|
+
|
|
11
|
+
VIDEO_EXT = {".mp4", ".mov", ".mkv", ".webm", ".avi", ".m4v"}
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
def collect_sources(srcs) -> list:
|
|
15
|
+
files = []
|
|
16
|
+
for raw in srcs:
|
|
17
|
+
p = Path(raw).expanduser()
|
|
18
|
+
if p.is_dir():
|
|
19
|
+
files += [f for f in sorted(p.rglob("*")) if f.suffix.lower() in VIDEO_EXT]
|
|
20
|
+
elif p.suffix.lower() in VIDEO_EXT:
|
|
21
|
+
files.append(media_path(p))
|
|
22
|
+
# 去重且保持顺序
|
|
23
|
+
seen, out = set(), []
|
|
24
|
+
for f in files:
|
|
25
|
+
key = str(f.resolve()).lower()
|
|
26
|
+
if key not in seen:
|
|
27
|
+
seen.add(key)
|
|
28
|
+
out.append(f.resolve())
|
|
29
|
+
return out
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
def pace_map(sections: list, default_k: int = 4, k_fast: int = 3, k_slow: int = 8) -> list:
|
|
33
|
+
"""按段落能量定每段的“几拍一切”:能量越高切得越密。
|
|
34
|
+
|
|
35
|
+
能量差小于 15% 时视为同一档,避免两段几乎一样响却切出 6 拍 vs 2 拍。
|
|
36
|
+
"""
|
|
37
|
+
if not sections or len(sections) < 2:
|
|
38
|
+
return []
|
|
39
|
+
energies = [float(s.get("energy") or 0) for s in sections]
|
|
40
|
+
lo, hi = min(energies), max(energies)
|
|
41
|
+
spread = (hi - lo) / max(hi, 1e-9)
|
|
42
|
+
uniform = spread < 0.15
|
|
43
|
+
out = []
|
|
44
|
+
for s, e in zip(sections, energies):
|
|
45
|
+
if uniform:
|
|
46
|
+
k = default_k
|
|
47
|
+
else:
|
|
48
|
+
norm = (e - lo) / (hi - lo) if hi > lo else 0.5
|
|
49
|
+
k = int(round(k_slow + (k_fast - k_slow) * norm))
|
|
50
|
+
out.append(
|
|
51
|
+
{
|
|
52
|
+
"start": s["start"],
|
|
53
|
+
"end": s["end"],
|
|
54
|
+
"energy": e,
|
|
55
|
+
"beats_per_cut": max(1, min(k, k_slow)),
|
|
56
|
+
}
|
|
57
|
+
)
|
|
58
|
+
return out
|
|
59
|
+
|
|
60
|
+
|
|
61
|
+
def cut_points(
|
|
62
|
+
beats: list,
|
|
63
|
+
duration: float,
|
|
64
|
+
beats_per_cut: int,
|
|
65
|
+
pmap: list | None = None,
|
|
66
|
+
fallback: float = 2.5,
|
|
67
|
+
) -> list:
|
|
68
|
+
if not beats:
|
|
69
|
+
n = max(1, int(duration // fallback))
|
|
70
|
+
return [round(i * fallback, 3) for i in range(n + 1)] + [round(duration, 3)]
|
|
71
|
+
cuts = [0.0]
|
|
72
|
+
if pmap:
|
|
73
|
+
for sec in pmap:
|
|
74
|
+
k = int(sec["beats_per_cut"])
|
|
75
|
+
bs = [b for b in beats if sec["start"] <= b < sec["end"]]
|
|
76
|
+
for i, b in enumerate(bs):
|
|
77
|
+
if (i + 1) % k == 0 and b > cuts[-1] + 0.4 and b < duration - 0.4:
|
|
78
|
+
cuts.append(round(b, 3))
|
|
79
|
+
else:
|
|
80
|
+
for i, b in enumerate(beats):
|
|
81
|
+
if (i + 1) % max(1, beats_per_cut) == 0 and b > cuts[-1] + 0.4 and b < duration - 0.4:
|
|
82
|
+
cuts.append(round(b, 3))
|
|
83
|
+
cuts.append(round(duration, 3))
|
|
84
|
+
return cuts
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
def build_pool(files: list, limit: int, min_shot: float, progress) -> list:
|
|
88
|
+
from .shots import detect_shots
|
|
89
|
+
|
|
90
|
+
pool = []
|
|
91
|
+
for idx, f in enumerate(files[:limit], 1):
|
|
92
|
+
try:
|
|
93
|
+
shots = detect_shots(f)
|
|
94
|
+
except Exception as exc:
|
|
95
|
+
progress(f"跳过 {f.name}: {exc}")
|
|
96
|
+
continue
|
|
97
|
+
kept = [s for s in shots if s["dur"] >= min_shot]
|
|
98
|
+
progress(f"[{idx}/{min(len(files), limit)}] {f.name}: {len(shots)} 镜头,{len(kept)} 个可用")
|
|
99
|
+
for s in kept:
|
|
100
|
+
pool.append({"src": str(f), "start": s["start"], "end": s["end"], "dur": s["dur"]})
|
|
101
|
+
return pool
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
def select_clips(cuts: list, pool: list, seed: int, progress) -> tuple:
|
|
105
|
+
if not pool:
|
|
106
|
+
raise RuntimeError("素材里没有可用镜头:检查 --src 路径或降低 --min-shot")
|
|
107
|
+
rng = random.Random(seed)
|
|
108
|
+
clips, used, last_key = [], set(), None
|
|
109
|
+
clamped, reused = 0, 0
|
|
110
|
+
cursor = 0.0
|
|
111
|
+
for cut in cuts[1:]:
|
|
112
|
+
need = cut - cursor
|
|
113
|
+
if need < 0.4:
|
|
114
|
+
continue
|
|
115
|
+
cands = [s for s in pool if s["dur"] >= need - 1e-3]
|
|
116
|
+
if not cands:
|
|
117
|
+
cands = [max(pool, key=lambda s: s["dur"])]
|
|
118
|
+
clamped += 1
|
|
119
|
+
fresh = [s for s in cands if (s["src"], s["start"]) not in used]
|
|
120
|
+
avoid_last = [s for s in (fresh or cands) if (s["src"], s["start"]) != last_key]
|
|
121
|
+
pick_pool = avoid_last or fresh or cands
|
|
122
|
+
shot = rng.choice(pick_pool)
|
|
123
|
+
if shot["dur"] < need:
|
|
124
|
+
need = shot["dur"]
|
|
125
|
+
key = (shot["src"], shot["start"])
|
|
126
|
+
if key in used:
|
|
127
|
+
reused += 1
|
|
128
|
+
used.add(key)
|
|
129
|
+
last_key = key
|
|
130
|
+
hi = max(shot["start"], shot["end"] - need)
|
|
131
|
+
in_t = rng.uniform(shot["start"], hi)
|
|
132
|
+
clips.append(
|
|
133
|
+
{
|
|
134
|
+
"src": shot["src"],
|
|
135
|
+
"in": round(in_t, 3),
|
|
136
|
+
"out": round(in_t + need, 3),
|
|
137
|
+
}
|
|
138
|
+
)
|
|
139
|
+
cursor += need
|
|
140
|
+
progress(f"选了 {len(clips)} 段,复用 {reused} 次,素材不够被压短 {clamped} 次")
|
|
141
|
+
return clips, {"reused": reused, "clamped": clamped, "filled": round(cursor, 3)}
|
|
142
|
+
|
|
143
|
+
|
|
144
|
+
def generate(
|
|
145
|
+
music,
|
|
146
|
+
srcs,
|
|
147
|
+
out,
|
|
148
|
+
size=(1080, 1920),
|
|
149
|
+
fps: int = 30,
|
|
150
|
+
fit: str = "crop",
|
|
151
|
+
beats_per_cut: int = 4,
|
|
152
|
+
pace: str = "auto",
|
|
153
|
+
transition: str = "none",
|
|
154
|
+
transition_duration: float = 0.25,
|
|
155
|
+
max_sources: int = 12,
|
|
156
|
+
min_shot: float = 0.8,
|
|
157
|
+
seed: int = 7,
|
|
158
|
+
gain_db: float = -14.0,
|
|
159
|
+
plan_only: bool = False,
|
|
160
|
+
progress=None,
|
|
161
|
+
) -> dict:
|
|
162
|
+
note = progress or (lambda *_: None)
|
|
163
|
+
m_path = media_path(music)
|
|
164
|
+
note(f"分析音乐 {m_path.name}")
|
|
165
|
+
m = analyze(m_path, with_beats=True)
|
|
166
|
+
pmap = pace_map(m.get("sections") or [], default_k=beats_per_cut) if pace == "auto" else []
|
|
167
|
+
cuts = cut_points(m["beats"], m["duration"], beats_per_cut, pmap=pmap)
|
|
168
|
+
if pace == "auto" and len(cuts) < 3:
|
|
169
|
+
cuts = cut_points(m["beats"], m["duration"], beats_per_cut)
|
|
170
|
+
pmap = []
|
|
171
|
+
note("段落变化不足,退回固定拍数")
|
|
172
|
+
note(f"{m['tempo_bpm']} BPM,{len(m['beats'])} 拍,切点 {len(cuts) - 1} 个,节奏 {pace}")
|
|
173
|
+
|
|
174
|
+
files = collect_sources(srcs)
|
|
175
|
+
note(f"素材 {len(files)} 个文件,本次扫描前 {min(len(files), max_sources)} 个")
|
|
176
|
+
pool = build_pool(files, max_sources, min_shot, note)
|
|
177
|
+
clips, stats = select_clips(cuts, pool, seed, note)
|
|
178
|
+
|
|
179
|
+
plan = {
|
|
180
|
+
"size": list(size),
|
|
181
|
+
"fps": fps,
|
|
182
|
+
"fit": fit,
|
|
183
|
+
"crf": 20,
|
|
184
|
+
"clips": clips,
|
|
185
|
+
"music": {"src": str(m_path), "gain_db": gain_db, "loop": True},
|
|
186
|
+
"keep_audio": False,
|
|
187
|
+
}
|
|
188
|
+
if transition and transition != "none":
|
|
189
|
+
plan["transition"] = {"type": transition, "duration": transition_duration}
|
|
190
|
+
out_path = Path(out).resolve()
|
|
191
|
+
ensure_dir(out_path.parent)
|
|
192
|
+
plan_path = out_path.with_suffix(out_path.suffix + ".plan.json")
|
|
193
|
+
plan_path.write_text(json.dumps(plan, ensure_ascii=False, indent=2), encoding="utf-8")
|
|
194
|
+
note(f"计划已写入 {plan_path}")
|
|
195
|
+
|
|
196
|
+
result = {
|
|
197
|
+
"out": None,
|
|
198
|
+
"plan": str(plan_path),
|
|
199
|
+
"duration": m["duration"],
|
|
200
|
+
"tempo_bpm": m["tempo_bpm"],
|
|
201
|
+
"beats": len(m["beats"]),
|
|
202
|
+
"cuts": cuts,
|
|
203
|
+
"segments": len(clips),
|
|
204
|
+
"sources": len(files),
|
|
205
|
+
"pool_shots": len(pool),
|
|
206
|
+
"seed": seed,
|
|
207
|
+
"stats": stats,
|
|
208
|
+
"pace": {"mode": pace, "sections": pmap},
|
|
209
|
+
"transition": transition if transition != "none" else None,
|
|
210
|
+
}
|
|
211
|
+
if plan_only:
|
|
212
|
+
return result
|
|
213
|
+
|
|
214
|
+
note("开始渲染")
|
|
215
|
+
rendered = render(plan, out_path)
|
|
216
|
+
result.update({"out": rendered["out"], "duration": rendered["duration"]})
|
|
217
|
+
return result
|