Async: Call the cloud audio API with a prompt about the given audio file. Audio is converted to MP3 once (not retried); the API call is retr

根据选择的方法进行分析

批量分析场景目录并返回统计信息。

综合相似度计算 注意：新增了 visual_vec 参数，直接传入预计算好的光影向量

Calculate the start and end frame indices based on the given time range. Args: ele (dict): A dictionary containing optional 'video_s

Generate caption for an audio file using Madmom keypoints for segmentation. This function produces a two-level hierarchical structure: -

Estimate token count. Rough heuristic: ~3-4 chars per token for English.

Extract mono 16k MP3 from video using ffmpeg.

Generate audio captions for multiple audio files via concurrent cloud API calls. Args: audio_paths: List of paths to audio files

Generate overall analysis for the entire audio file via cloud API. Args: audio_path: Path to the audio file temperature: Sam

Get video width and height using ffprobe.

从 decord VideoReader 加载场景帧（内存读取，无磁盘 I/O） frame_range: [start_frame, end_frame]，单位为采样帧空间索引 frame_indices: decode_video_to_frames 返回的原始视频帧索

Load scene_caption.scene_summary from all scene JSON files in a folder. Skips non-usable scenes and scenes with importance_score < 3. Return

Merge multiple sections by index Args: sections: List of all sections indices: List of section indices to merge Returns

自适应阈值的NMS，阈值 = 最大值 * adaptive_ratio

窗口NMS：每个时间窗口内保留top-k个点

按类型归一化关键点强度，使不同类型的关键点可以公平比较 每种类型使用其主要特征值进行归一化： - Downbeat: 使用 activation (DBN激活值) 归一化后所有类型的强度都在 0 ~ 1 范围内 Args: keypoi

Robustly parse a JSON string, stripping Markdown code fences if present.

Parse an SRT file into a list of subtitle dicts. Each dict has keys: start_sec, end_sec, speaker (or None), text.

Parse structure proposal JSON from LLM output. Expected format:: { "overall_theme": "...", "narrative_logic": ".

Encode a PIL Image to a base64 JPEG string.

Read and process audio info without using librosa. Support dict keys: type = audio - audio - audio_start - audio_end t

Process video and generate captions (Step 1: clip captioning, Step 2: scene merge). Scene video analysis (Step 3) is handled separately by th

Resample audio to target sample rate using scipy. Args: audio: numpy array of audio samples orig_sr: Original sample rat

Round a dimension to the nearest positive even integer for video encoding.

Run the screenwriter pipeline to generate a shot plan.

Extract a segment from an audio file. Args: audio_path: Path to the source audio file start_time: Start time in seconds

Select ONE opening dialogue clip for the whole video (target ~10s).

Remove leading speaker tag like [Mia Dolan] from subtitle text.

Write SRT file from sentence_info list. Each item: {'text': str, 'timestamp': [[word, start_ms, end_ms], ...], 'speaker': str|None}

Build the litellm message list for one clip (no LLM call).

Encode one audio segment to base64 and return (message_list, segment_start_time, tmp_path).

Build per-line absolute and relative timing records.

调用 VLM

Check if video file has audio track.

Compute non-silent intervals in seconds. This is a lightweight guardrail to prevent keypoints from appearing in silent regions where onset mo

[修改] 使用缓存计算标签集合相似度

Return a dict of losers keyed by (sec_idx, shot_idx) with guidance text.

Compute fuzzy similarity between two normalized subtitle lines.

Ensure the given audio path is a WAV file readable by aubio. Aubio's `source_wavread` requires a RIFF/WAV container. Gradio uploads are often

Evaluate a raw VLM detection result and apply size/role rules.

Extract the first balanced {...} JSON object from mixed text.

Reuse video_caption.parse_srt_to_dict() and only do range filtering + formatting here.

Generate cache key based on audio file and parameters

Return GPU context if decord was compiled with CUDA support, otherwise CPU. The check is cached after the first call via _create_decord_reader, w

判断是否为桥接镜头

Pre-load audio as waveform dict to bypass torchcodec/AudioDecoder incompatibility in pyannote 4.x. Returns {'waveform': (1, time) torch.Tenso

Load cached result from disk

Load progress from existing output file to support resume functionality. Returns a set of completed (section_idx, shot_idx) tuples.

Match model-selected lines back to original SRT subtitle entries.

Merge consecutive segments from the same speaker. Args: segments: Original segment list max_gap: Maximum time gap allowed fo

将极短 scene 合并到相邻 scene：首段并后，其余并前。

Parse JSON response for audio section selection.

Parse LLM output into a JSON object with light normalization.

Parse SRT response and adjust timestamps.

Parse shot ranges in the format: [shot: HH:MM:SS to HH:MM:SS].

Work around a Gradio CheckboxGroup edge-case. Some Gradio versions may send `null` for an empty CheckboxGroup selection, which becomes `None`

Persist target output length by materializing the derived min/max config values.

Persist target shot length by materializing the derived min/max config values.

双阶段后处理： 1. Forward Pass: 处理短的 Establishment (前奏合并) 2. Backward Pass: 处理短的 Progression (碎片回收)

[新增] 预计算阶段： 1. 提取所有出现的唯一环境标签，批量编码。 2. 提取所有 Shot 的光影色调描述，批量编码。

Read key=value pairs from config.py as strings.

Resolve video scene path from a raw video path.

Split video into segments and run scenedetect in parallel across processes. Each segment starts decoding `warmup_frames` before the logical segme

Save result to disk cache

拆分超过最大时长的场景 Args: scenes: 场景列表，每个场景是一个 shot 列表 max_duration_secs: 单个场景允许的最大时长（秒），默认5分钟 Returns:

Transcribe audio via LiteLLM using a multimodal model (e.g. Gemini). Audio is automatically split into chunks no larger than max_segment_siz

Run whisper.cpp ASR via pywhispercpp with optional pyannote speaker diarization. Returns a dict with keys: - text: full transcription

Validate parsed shot plan structure.

Async: Call the cloud audio API concurrently for multiple audio files. Uses asyncio.gather with a semaphore to limit concurrent requests.

Adjust clip start time if there's a scene cut within the clip. If a scene cut point exists within the clip (excluding the exact start/end),

Mel能量检测 + NMS