let5see/AI-Video
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update prompts and content generators

Browse Source
This commit is contained in:
张欣
2025-12-05 16:42:10 +08:00
parent 5739601afa
commit 1769b7faee
11 changed files with 518 additions and 260 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
pixelle_video/prompts/asset_script_generation.py
View File

@@ -17,37 +17,38 @@ For generating video scripts based on user-provided assets.
"""
ASSET_SCRIPT_GENERATION_PROMPT = """你是一位专业的视频脚本创作者。请基于用户提供的视频意图和可用素材，生成一个 {duration} 秒的视频脚本。
ASSET_SCRIPT_GENERATION_PROMPT = """You are a professional video script creator. Based on the user's video intent and available assets, generate a {duration}-second video script. Before doing so, you need to detect the user's input language - if it's English, then all copy must be in English. Strictly follow the user's input language type as the standard, ensuring consistent and corresponding copy!
## 需求信息
{title_section}- 视频意图：{intent}
- 目标时长：{duration} 秒
## Requirements
{title_section}- Video Intent: {intent}
- Target Duration: {duration} seconds
## 可用素材（请在输出中使用精确路径）
## Available Assets (use exact paths in output)
{assets_text}
## 创作指南
1. 根据目标时长决定需要多少个场景（通常每个场景 5-15 秒）
2. 为每个场景从可用素材中直接分配一个素材
3. 每个场景可以包含 1-3 句旁白
4. 尽量使用所有可用素材，但如有需要可以复用素材
5. 所有场景的总时长应约等于 {duration} 秒
## Creation Guidelines
1. Strictly output copy according to the user's input language type - if input is English, output must be English, and so on
2. Determine the number of scenes based on target duration (typically 5-15 seconds per scene)
3. Assign one asset from available assets to each scene
4. Each scene can contain 1-3 narration sentences
5. Try to use all available assets, but assets can be reused if needed
6. Total duration of all scenes should approximately equal {duration} seconds
{title_instruction}
## 语言一致性要求（非常重要）
- 旁白的语言必须与用户输入的视频意图保持一致
- 如果视频意图是中文，则旁白必须是中文
- 如果视频意图是英文，则旁白必须是英文
- 除非视频意图中明确指定了输出语言，否则严格遵循意图的原始语言
## Language Consistency Requirements (Strictly Enforce)
- Narration language must match the user's input video intent
- If video intent is in Chinese, narration must be in Chinese
- If video intent is in English, narration must be in English
- Unless the video intent explicitly specifies an output language, strictly follow the original language of the intent
## 输出要求
为每个场景提供：
- scene_number: 场景编号（从 1 开始）
- asset_path: 从可用素材列表中选择的精确路径
- narrations: 包含 1-3 句旁白的数组
- duration: 预估时长（秒）
## Output Requirements
Provide for each scene:
- scene_number: Scene number (starting from 1)
- asset_path: Exact path selected from available assets list
- narrations: Array containing 1-3 narration sentences
- duration: Estimated duration (seconds)
现在请开始生成视频脚本："""
Now please begin generating the video script:"""
def build_asset_script_prompt(
@@ -68,8 +69,8 @@ def build_asset_script_prompt(
Returns:
Formatted prompt
"""
title_section = f"- 视频标题：{title}\n" if title else ""
title_instruction = f"6. 旁白内容应与视频标题保持一致：{title}\n" if title else ""
title_section = f"- Video Title: {title}\n" if title else ""
title_instruction = f"6. Narration content should be consistent with the video title: {title}\n" if title else ""
return ASSET_SCRIPT_GENERATION_PROMPT.format(
duration=duration,

80
pixelle_video/prompts/content_narration.py
View File

@@ -17,61 +17,63 @@ For extracting/refining narrations from user-provided content.
"""
CONTENT_NARRATION_PROMPT = """# 角色定位
你是一位专业的内容提炼专家，擅长从用户提供的内容中提取核心要点，并转化成适合短视频的脚本。
CONTENT_NARRATION_PROMPT = """# Role Definition
Globally, you must strictly output copy in the corresponding language type according to the user's language type.
You are a professional content refinement expert, skilled at extracting core points from user-provided content and transforming them into scripts suitable for short videos.
# 核心任务
用户会提供一段内容（可能很长，也可能很短），你需要从中提炼出 {n_storyboard} 个视频分镜的旁白（用于TTS生成视频音频）。
# Core Task
The user will provide content (which may be long or short), and you need to extract narrations for {n_storyboard} video storyboards (for TTS to generate video audio).
# 用户提供的内容
# User-Provided Content
{content}
# 输出要求
# Output Requirements
## 旁白规范
- 用途定位：用于TTS生成短视频音频
- 字数限制：严格控制在{min_words}~{max_words}个字（最低不少于{min_words}字）
- 结尾格式：结尾不要使用标点符号
- 提炼策略：
* 如果用户内容较长：提取{n_storyboard}个核心要点，去除冗余信息
* 如果用户内容较短：在保留核心观点的基础上适当扩展，增加例子或解释
* 如果用户内容刚好：优化表达，使其更适合口播
- 风格要求：保持用户内容的核心观点，但用更口语化、适合TTS的方式表达
- 开场建议：第一个分镜可以用提问或场景引入，吸引观众注意
- 核心内容：中间分镜展开用户内容的核心要点
- 结尾建议：最后一个分镜给出总结或启发
- 情绪与语气：温和、真诚、自然，像在跟朋友分享观点
- 禁止项：不出现网址、表情符号、数字编号、不说空话套话
- 字数检查：生成后必须自我验证每段不少于{min_words}个字
## Narration Specifications
- Language consistency requirement: Strictly output copy according to the user's input language type - if input is English, output must be English, and so on
- Purpose: For TTS to generate short video audio
- Word count limit: Strictly control to {min_words}~{max_words} words (minimum not less than {min_words} words)
- Ending format: Do not use punctuation at the end
- Refinement strategy:
* If user content is long: Extract {n_storyboard} core points, remove redundant information
* If user content is short: Appropriately expand while retaining core viewpoints, add examples or explanations
* If user content is just right: Optimize expression to make it more suitable for voice narration
- Style requirement: Maintain the core viewpoint of user content, but express it in a more colloquial way suitable for TTS
- Opening suggestion: The first storyboard can use a question or scene introduction to attract audience attention
- Core content: Middle storyboards expand on the core points of user content
- Ending suggestion: The last storyboard provides a summary or inspiration
- Emotion and tone: Gentle, sincere, natural, like sharing viewpoints with a friend
- Prohibitions: No URLs, emojis, numeric numbering, no empty talk or clichés
- Word count check: After generation, must self-verify that each segment is not less than {min_words} words
## 分镜连贯性要求
- {n_storyboard} 个分镜应基于用户内容的核心观点展开，形成完整表达
- 保持逻辑连贯，自然过渡
- 每个分镜像同一个人在讲述，语气一致
- 确保提炼的内容忠于用户原意，但更适合短视频呈现
## Storyboard Coherence Requirements
- {n_storyboard} storyboards should expand based on the core viewpoint of user content, forming a complete expression
- Maintain logical coherence and natural transitions
- Each storyboard should sound like the same person narrating, with consistent tone
- Ensure the refined content is faithful to the user's original meaning, but more suitable for short video presentation
# 输出格式
严格按照以下JSON格式输出，不要添加任何额外的文字说明：
# Output Format
Strictly output in the following JSON format, do not add any additional text explanations:
```json
{{
"narrations": [
"第一段{min_words}~{max_words}字的旁白",
"第二段{min_words}~{max_words}字的旁白",
"第三段{min_words}~{max_words}字的旁白"
"First {min_words}~{max_words} word narration",
"Second {min_words}~{max_words} word narration",
"Third {min_words}~{max_words} word narration"
]
}}
```
# 重要提醒
1. 只输出JSON格式内容，不要添加任何解释说明
2. 确保JSON格式严格正确，可以被程序直接解析
3. 旁白必须严格控制在{min_words}~{max_words}字之间
4. 必须输出恰好 {n_storyboard} 个分镜的旁白
5. 内容要忠于用户原意，但优化为更适合口播的表达
6. 输出格式为 {{"narrations": [旁白数组]}} 的JSON对象
# Important Reminders
1. Only output JSON format content, do not add any explanations
2. Ensure JSON format is strictly correct and can be directly parsed by the program
3. Narrations must be strictly controlled between {min_words}~{max_words} words
4. Must output exactly {n_storyboard} storyboard narrations
5. Content must be faithful to the user's original meaning, but optimized for voice narration expression
6. Output format is {{"narrations": [narration array]}} JSON object
现在，请从上述内容中提炼出 {n_storyboard} 个分镜的旁白。只输出JSON，不要其他内容。
Now, please extract {n_storyboard} storyboard narrations from the above content. Only output JSON, no other content.
"""

102
pixelle_video/prompts/image_generation.py
View File

@@ -25,21 +25,21 @@ from typing import List, Optional
IMAGE_STYLE_PRESETS = {
"stick_figure": {
"name": "火柴人简笔画",
"name": "Stick Figure Sketch",
"description": "stick figure style sketch, black and white lines, pure white background, minimalist hand-drawn feel",
"use_case": "通用场景，简单直观"
"use_case": "General scenes, simple and intuitive"
},
"minimal": {
"name": "极简抽象",
"name": "Minimalist Abstract",
"description": "minimalist abstract art, geometric shapes, clean composition, modern design, soft pastel colors",
"use_case": "现代感、艺术感"
"use_case": "Modern, artistic feel"
},
"concept": {
"name": "概念化视觉",
"name": "Conceptual Visual",
"description": "conceptual visual metaphors, symbolic elements, thought-provoking imagery, artistic interpretation",
"use_case": "深度内容、哲学思考"
"use_case": "Deep content, philosophical thinking"
},
}
@@ -47,52 +47,52 @@ IMAGE_STYLE_PRESETS = {
DEFAULT_IMAGE_STYLE = "stick_figure"
IMAGE_PROMPT_GENERATION_PROMPT = """# 角色定位
你是一个专业的视觉创意设计师，擅长为视频脚本创作富有表现力和象征性的图像提示词，将抽象概念转化为具象的视觉画面。
IMAGE_PROMPT_GENERATION_PROMPT = """# Role Definition
You are a professional visual creative designer, skilled at creating expressive and symbolic image prompts for video scripts, transforming abstract concepts into concrete visual scenes.
# 核心任务
基于已有的视频脚本，为每个分镜的"旁白内容"创作对应的**英文**图像提示词，确保视觉画面与叙述内容完美配合，增强观众的理解和记忆。
# Core Task
Based on the existing video script, create corresponding **English** image prompts for each storyboard's "narration content", ensuring visual scenes perfectly match the narrative content and enhance audience understanding and memory.
**重要：输入包含 {narrations_count} 个旁白，你必须为每个旁白都生成一个对应的图像提示词，总共输出 {narrations_count} 个图像提示词。**
**Important: The input contains {narrations_count} narrations. You must generate one corresponding image prompt for each narration, totaling {narrations_count} image prompts.**
# 输入内容
# Input Content
{narrations_json}
# 输出要求
# Output Requirements
## 图像提示词规范
- 语言：**必须使用英文**（用于 AI 图像生成模型）
- 描述结构：scene + character action + emotion + symbolic elements
- 描述长度：确保描述清晰完整且富有创意（建议 50-100 个英文单词）
## Image Prompt Specifications
- Language: **Must use English** (for AI image generation models)
- Description structure: scene + character action + emotion + symbolic elements
- Description length: Ensure clear, complete, and creative descriptions (recommended 50-100 English words)
## 视觉创意要求
- 每个图像都要准确反映对应旁白的具体内容和情感
- 使用象征手法将抽象概念视觉化（如用路径代表人生选择，用锁链代表束缚等）
- 画面要表现出丰富的情感和动作，增强视觉冲击力
- 通过构图和元素安排突出主题，避免过于直白的表现方式
## Visual Creative Requirements
- Each image must accurately reflect the specific content and emotion of the corresponding narration
- Use symbolic techniques to visualize abstract concepts (e.g., use paths to represent life choices, chains to represent constraints, etc.)
- Scenes should express rich emotions and actions to enhance visual impact
- Highlight themes through composition and element arrangement, avoid overly literal representations
## 关键英文词汇参考
- 象征元素：symbolic elements
- 表情：expression / facial expression
- 动作：action / gesture / movement
- 场景：scene / setting
- 氛围：atmosphere / mood
## Key English Vocabulary Reference
- Symbolic elements: symbolic elements
- Expression: expression / facial expression
- Action: action / gesture / movement
- Scene: scene / setting
- Atmosphere: atmosphere / mood
## 视觉与文案配合原则
- 图像要服务于文案，成为文案内容的视觉延伸
- 避免与文案内容无关或矛盾的视觉元素
- 选择最能增强文案说服力的视觉表现方式
- 确保观众能通过图像快速理解文案的核心观点
## Visual and Copy Coordination Principles
- Images should serve the copy, becoming a visual extension of the copy content
- Avoid visual elements unrelated to or contradicting the copy content
- Choose visual presentation methods that best enhance the persuasiveness of the copy
- Ensure the audience can quickly understand the core viewpoint of the copy through images
## 创意指导
1. **现象描述类文案**：用直观的场景表现社会现象
2. **原因分析类文案**：用因果关系的视觉比喻表现内在逻辑
3. **影响论证类文案**：用后果场景或对比手法表现影响程度
4. **深入探讨类文案**：用抽象概念的具象化表现深刻思考
5. **结论启发类文案**：用开放式场景或指引性元素表现启发性
## Creative Guidance
1. **Phenomenon Description Copy**: Use intuitive scenes to represent social phenomena
2. **Cause Analysis Copy**: Use visual metaphors of cause-and-effect relationships to represent internal logic
3. **Impact Argumentation Copy**: Use consequence scenes or contrast techniques to represent the degree of impact
4. **In-depth Discussion Copy**: Use concretization of abstract concepts to represent deep thinking
5. **Conclusion Inspiration Copy**: Use open-ended scenes or guiding elements to represent inspiration
# 输出格式
严格按照以下JSON格式输出，**图像提示词必须是英文**：
# Output Format
Strictly output in the following JSON format, **image prompts must be in English**:
```json
{{
@@ -103,17 +103,17 @@ IMAGE_PROMPT_GENERATION_PROMPT = """# 角色定位
}}
```
# 重要提醒
1. 只输出JSON格式内容，不要添加任何解释说明
2. 确保JSON格式严格正确，可以被程序直接解析
3. 输入是 {{"narrations": [旁白数组]}} 格式，输出是 {{"image_prompts": [图像提示词数组]}} 格式
4. **输出的image_prompts数组必须恰好包含 {narrations_count} 个元素，与输入的narrations数组一一对应**
5. **图像提示词必须使用英文**（for AI image generation models）
6. 图像提示词必须准确反映对应旁白的具体内容和情感
7. 每个图像都要有创意性和视觉冲击力，避免千篇一律
8. 确保视觉画面能增强文案的说服力和观众的理解度
# Important Reminders
1. Only output JSON format content, do not add any explanations
2. Ensure JSON format is strictly correct and can be directly parsed by the program
3. Input is {{"narrations": [narration array]}} format, output is {{"image_prompts": [image prompt array]}} format
4. **The output image_prompts array must contain exactly {narrations_count} elements, corresponding one-to-one with the input narrations array**
5. **Image prompts must use English** (for AI image generation models)
6. Image prompts must accurately reflect the specific content and emotion of the corresponding narration
7. Each image must be creative and visually impactful, avoid being monotonous
8. Ensure visual scenes can enhance the persuasiveness of the copy and audience understanding
现在，请为上述 {narrations_count} 个旁白创作对应的 {narrations_count} 个**英文**图像提示词。只输出JSON，不要其他内容。
Now, please create {narrations_count} corresponding **English** image prompts for the above {narrations_count} narrations. Only output JSON, no other content.
"""

56
pixelle_video/prompts/title_generation.py
View File

@@ -17,35 +17,69 @@ For generating video title from content.
"""
TITLE_GENERATION_PROMPT = """Please generate a short, attractive title (within 10 characters) for the following content.
TITLE_GENERATION_PROMPT = """Please generate a short, attractive title for the following content.
Content:
{content}
Requirements:
1. Brief and concise, within 10 characters
2. Accurately summarize the core content
3. Attractive, suitable as a video title
4. Output only the title text, no other content
1. **Language Consistency (CRITICAL)**: The title MUST be in the same language as the input content
- If the input content is in English, the title MUST be in English
- If the input content is in Chinese, the title MUST be in Chinese
- Strictly follow the language of the input content
2. **Character Limit (CRITICAL)**: The title MUST NOT exceed {max_length} characters
- Count every character including spaces
- The title must be complete and meaningful within this limit
- Do NOT generate a title that would need to be cut off
3. **Core Message (CRITICAL)**: The title MUST capture the MAIN POINT of the content
- Identify the central theme or key message
- Don't focus on just one aspect if the content has multiple important points
- Ensure the title accurately represents what the content is about
4. **No Punctuation at End**: Do NOT include any punctuation marks at the end of the title
- No period (.), comma (,), exclamation mark (!), question mark (?), etc.
- The title should end with a word or number, not punctuation
5. **Completeness**: Ensure the title is a complete, meaningful phrase
- Do not cut off in the middle of a word or number
- Do not create incomplete phrases like "Rise Early for" or "How to Make"
- Use abbreviations or shorter words if needed to fit the limit
6. **Abbreviation Examples** (use when needed to fit character limit):
- For English:
* "10,000" → "10K"
* "per month" → "monthly" or "a month"
* "early to bed and early to rise" → "Sleep Early" or "Early Habits"
* "makes you healthy" → "for Health" or "Stay Healthy"
- For Chinese:
* "10,000元" → "万元" or "1万"
* "每个月" → "月入" or "月收"
7. Accurately summarize the core content
8. Attractive and engaging, suitable as a video title
9. Output only the title text, no quotes, no explanations
Title:"""
def build_title_generation_prompt(content: str, max_length: int = 500) -> str:
def build_title_generation_prompt(content: str, max_length: int = 15) -> str:
"""
Build title generation prompt
Args:
content: Content to generate title from
max_length: Maximum content length to use (default 500 chars)
max_length: Maximum title length in characters (default: 15)
Returns:
Formatted prompt
Formatted prompt with character limit
"""
# Take first max_length chars to avoid overly long prompts
content_preview = content[:max_length]
# Take first 500 chars to avoid overly long prompts
content_preview = content[:500]
return TITLE_GENERATION_PROMPT.format(
content=content_preview
content=content_preview,
max_length=max_length
)

163
pixelle_video/prompts/topic_narration.py
View File

@@ -17,110 +17,117 @@ For generating narrations from a topic/theme.
"""
TOPIC_NARRATION_PROMPT = """# 角色定位
你是一位专业的内容创作专家，擅长将话题扩展成引人入胜的短视频脚本，用深入浅出的方式讲解观点，帮助观众理解复杂概念。
TOPIC_NARRATION_PROMPT = """# Role Definition
You are a professional content creation expert, skilled at expanding topics into engaging short video scripts, explaining viewpoints in an accessible way to help audiences understand complex concepts.
Globally, you must strictly output copy in the corresponding language type according to the user's language type.
# 核心任务
用户会输入一个话题或主题，你需要为这个话题或主题进行创作 {n_storyboard} 个视频分镜，每个分镜包含"旁白（用于TTS生成视频讲解音频）"，像在跟朋友聊天一样，自然、有价值、引发共鸣。
# Core Task
The user will input a topic or theme. You need to create {n_storyboard} video storyboards for this topic or theme. Each storyboard contains "narration (for TTS to generate video explanation audio)", naturally and valuably, like chatting with a friend, to resonate with the audience.
- Language consistency requirement: Strictly output copy according to the user's input language type - if input is English, output must be English, and so on
# 输入话题
# Input Topic
{topic}
# 输出要求
# Output Requirements
## 旁白规范
- 输出语言要求：严格按照用户输入的话题或主题的语种输出，如：用户输入的是英文，则输出的文案必须为英文，中文也是一样。
- 用途定位：用于TTS生成短视频音频，通俗易懂地讲解话题
- 字数限制：严格控制在{min_words}~{max_words}个字（最低不少于{min_words}字）
- 结尾格式：每段旁白的结尾不要使用标点符号，若旁白中出现断句读法必须使用中文标点（，。？！……：“”）来表达语气和停顿，自动判断并插入合适的标点符号，保留自然口语节奏（比如“对吗？不对。”要有停顿和语气转折）
- 内容要求：围绕话题展开，每个分镜传递一个有价值的观点或洞察
- 风格要求：像跟朋友聊天一样，通俗、真诚、有启发性，避免学术化和生硬的表达，拒绝套路化和模板化的表达
- 情绪与语气：温和、真诚、有热情，像一个有见解的朋友在分享思考
- 可适当引用权威内容，不强制每次输出都要有引用出现，根据用户传入的标题或内容参考判断是否需要有相关引用：
若为科学/健康类，可引用《自然》《柳叶刀》、哈佛研究、神经科学发现等；
若为心理/哲学类，可引用荣格、尼采、庄子、曾仕强、卡巴金等人的观点或语录；
若为国学/佛道类，可引用《道德经》《金刚经》《黄帝内经》等经典原文或释义；
若为文学/历史类，可引用鲁迅、苏轼、《史记》、《人类简史》等；
若为时尚/生活方式类，可引用色彩心理学、形象管理理论、行为经济学等。
根据上述举例，若有其他类型的方向和赛道也可检索引用相关书籍，但也要遵循不强制引用的要求。
## Narration Specifications
- Output language requirement: Strictly output according to the language of the user's input topic or theme. For example: if the user's input is in English, the output copy must be in English, same for Chinese.
- Purpose: For TTS to generate short video audio, explaining topics in an accessible way
- Word count limit: Strictly control to {min_words}~{max_words} words (minimum not less than {min_words} words)
- Ending format: Do not use punctuation at the end of each narration. If there are sentence breaks in the narration, Chinese punctuation (,。?!……:"") must be used to express tone and pauses. Automatically determine and insert appropriate punctuation to maintain natural spoken rhythm (e.g., "Right? Wrong." should have pauses and tonal shifts)
- Content requirement: Expand around the topic, each storyboard conveys a valuable viewpoint or insight
- Style requirement: Like chatting with a friend, accessible, sincere, inspiring, avoid academic and stiff expressions, reject formulaic and template expressions
- Emotion and tone: Gentle, sincere, enthusiastic, like a friend with insights sharing thoughts
- Can appropriately cite authoritative content, not mandatory for every output, determine based on the user's input title or content reference whether relevant citations are needed:
For science/health topics, can cite Nature, The Lancet, Harvard research, neuroscience findings, etc.;
For psychology/philosophy topics, can cite viewpoints or quotes from Jung, Nietzsche, Zhuangzi, Zeng Shiqiang, Kabat-Zinn, etc.;
For Chinese studies/Buddhism/Taoism topics, can cite original texts or interpretations from Tao Te Ching, Diamond Sutra, Yellow Emperor's Inner Canon, etc.;
For literature/history topics, can cite Lu Xun, Su Shi, Records of the Grand Historian, Sapiens, etc.;
For fashion/lifestyle topics, can cite color psychology, image management theory, behavioral economics, etc.
Based on the above examples, if there are other types of directions and tracks, relevant books can also be searched and cited, but must also follow the non-mandatory citation requirement.
若有引用需自然融入，不生硬堆砌，不虚构出处。
If there are citations, integrate them naturally, do not pile them up stiffly, do not fabricate sources.
## 开头多样性要求（最重要）
【核心原则】每个分镜的开头必须根据内容本身自然表达，拒绝任何形式的固定套路和模板化表达。
## Opening Diversity Requirements (Most Important)
[Core Principle] The opening of each storyboard must be expressed naturally based on the content itself, rejecting any form of fixed routines and template expressions.
【表达方式灵活性】
根据话题内容，可以采用陈述、场景、感叹、观点、问句、对比、故事等多种表达方式，但务必做到：
- 每个分镜根据要表达的具体内容选择最自然的开头
- 绝不形成任何规律性的句式模式
- 不要让任何一个词或短语成为"习惯性开头"
[Expression Flexibility]
Based on the topic content, various expression methods such as statements, scenes, exclamations, viewpoints, questions, contrasts, stories, etc. can be used, but must achieve:
- Each storyboard chooses the most natural opening based on the specific content to be expressed
- Never form any regular sentence pattern
- Do not let any word or phrase become a "habitual opening"
【严禁固定模式】
❌ 绝对禁止以下行为：
- 形成"第N句总用X开头"的任何规律
- 多次重复使用同一个连接词或句式作为开头
- 按照某种隐藏的模板顺序来组织分镜
[Strictly Prohibit Fixed Patterns]
❌ Absolutely prohibit the following behaviors:
- Forming any pattern of "the Nth sentence always starts with X"
- Repeatedly using the same conjunction or sentence pattern as an opening
- Organizing storyboards according to some hidden template order
【特别强调】
- 第一个分镜的开头要完全根据话题内容自然选择，不要有任何固定词汇倾向
- 整组旁白中，如果某个词（如"有时候"、"其实"、"你有没有"）出现超过1次作为开头，就是失败的创作
- 要像真人说话一样自然流畅，而不是套用任何句式模板
[Special Emphasis]
## Language Consistency Requirements (Strictly Enforce)
- Narration language must match the user's input video intent
- If video intent is in Chinese, narration must be in Chinese
- If video intent is in English, narration must be in English
- Unless the video intent explicitly specifies an output language, strictly follow the original language of the intent
- The opening of the first storyboard should be completely naturally chosen based on the topic content, without any fixed vocabulary tendency
- In the entire set of narrations, if any word (such as "sometimes", "actually", "have you ever") appears more than once as an opening, it is a failed creation
- Should be as natural and fluent as a real person speaking, not applying any sentence pattern template
## 自然表达要求
- 内容应该像真人在自然交流，而不是按照模板填空
- 每个分镜的开头要根据内容本身选择最合适的表达方式
- 同一个词作为开头在整个旁白中最多只能出现1次
- 优先用观点、场景、故事来串联内容，避免依赖连接词开头
## Natural Expression Requirements
- Content should be like real people communicating naturally, not filling in templates
- The opening of each storyboard should choose the most appropriate expression method based on the content itself
- The same word can appear as an opening at most once in the entire narration
- Prioritize using viewpoints, scenes, stories to connect content, avoid relying on conjunctions as openings
## 内容结构建议
- 开场方式：可以用场景、故事、观点、现象等多种方式引入，不固定套路
- 核心内容：中间分镜展开核心观点，用生活化的例子帮助理解
- 结尾方式：最后分镜给出行动建议或启发，让观众有收获感
- 整体逻辑：遵循"引发共鸣 → 提出观点 → 深入讲解 → 给出启发"的叙述逻辑
## Content Structure Suggestions
- Opening method: Can use scenes, stories, viewpoints, phenomena, and other methods to introduce, no fixed routine
- Core content: Middle storyboards expand core viewpoints, use life examples to help understanding
- Ending method: Last storyboard provides action suggestions or inspiration, giving the audience a sense of gain
- Overall logic: Follow the narrative logic of "resonate → propose viewpoint → in-depth explanation → provide inspiration"
## 其他规范
- 禁止项：不出现网址、表情符号、数字编号、不说空话套话、不过度煽情
- 字数检查：生成后必须自我验证不少于{min_words}个字，如不足则补充具体观点或例子
## Other Specifications
- Prohibitions: No URLs, emojis, numeric numbering, no empty talk or clichés, no excessive sentimentality
- Word count check: After generation, must self-verify not less than {min_words} words. If insufficient, supplement with specific viewpoints or examples
## 分镜连贯性要求
- {n_storyboard} 个分镜应围绕话题展开，形成完整的观点表达
- 遵循"吸引注意 → 提出观点 → 深入讲解 → 给出启发"的叙述逻辑
- 每个分镜像同一个人在连贯分享观点，语气一致、自然流畅
- 通过观点的递进自然过渡，形成完整的论述脉络
- 确保内容有价值、有启发，让观众觉得"这个视频值得看"
## Storyboard Coherence Requirements
- {n_storyboard} storyboards should expand around the topic, forming a complete viewpoint expression
- Follow the narrative logic of "attract attention → propose viewpoint → in-depth explanation → provide inspiration"
- Each storyboard should sound like the same person continuously sharing viewpoints, with consistent and natural tone
- Naturally transition through the progression of viewpoints, forming a complete argumentative thread
- Ensure content is valuable and inspiring, making the audience feel "this video is worth watching"
# 输出格式
严格按照以下JSON格式输出，不要添加任何额外的文字说明：
# Output Format
Strictly output in the following JSON format, do not add any additional text explanations:
```json
{{
"narrations": [
"第一段旁白内容",
"第二段旁白内容",
"第三段旁白内容"
"First narration content",
"Second narration content",
"Third narration content"
]
}}
```
# 重要提醒
1. 只输出JSON格式内容，不要添加任何解释说明
2. 确保JSON格式严格正确，可以被程序直接解析
3. 旁白必须严格控制在{min_words}~{max_words}字之间，用通俗易懂的语言
4. {n_storyboard} 个分镜要围绕话题展开，形成完整的观点表达
5. 每个分镜都要有价值，提供洞察，避免空洞的陈述
6. 输出格式为 {{"narrations": [旁白数组]}} 的JSON对象
# Important Reminders
1. Only output JSON format content, do not add any explanations
2. Ensure JSON format is strictly correct and can be directly parsed by the program
3. Narrations must be strictly controlled between {min_words}~{max_words} words, using accessible language
4. {n_storyboard} storyboards should expand around the topic, forming a complete viewpoint expression
5. Each storyboard must be valuable, providing insights, avoiding empty statements
6. Output format is {{"narrations": [narration array]}} JSON object
【多样性核心要求 - 必须严格执行】
7. 第一句旁白不要固定用某个词开头，每次创作都要根据话题内容自然选择不同的开头
8. 同一个词（如"有时候"、"你有没有"、"其实"、"想象一下"等）在所有旁白中作为开头最多只能出现1次
9. 不要形成任何隐藏的句式规律，每个分镜的开头要真正做到独立思考、自然表达
10. 检查你的输出：如果发现有任何词作为开头重复出现2次或以上，必须修改
11. 输出语言要求：严格按照用户输入的话题或主题的语种输出，如：用户输入的是英文，则输出的文案必须为英文，中文也是一样。
[Diversity Core Requirements - Must Strictly Execute]
7. The first narration should not use a fixed word as an opening. Each creation should naturally choose different openings based on the topic content
8. The same word (such as "sometimes", "have you ever", "actually", "imagine") can appear as an opening at most once in all narrations
9. Do not form any hidden sentence pattern rules. The opening of each storyboard should truly be independently thought out and naturally expressed
10. Check your output: if any word appears as an opening 2 or more times, it must be modified
11. Output language requirement: Strictly output according to the language of the user's input topic or theme. For example: if the user's input is in English, the output copy must be in English, same for Chinese.
现在，请为话题创作 {n_storyboard} 个分镜的旁白。
⚠️ 特别注意：写完后自查所有分镜的开头，确保没有重复使用同一个词或短语作为开头。
只输出JSON，不要其他内容。
Now, please create narrations for {n_storyboard} storyboards for the topic.
⚠️ Special note: After writing, self-check the openings of all storyboards to ensure no repeated use of the same word or phrase as an opening.
Only output JSON, no other content.
"""

106
pixelle_video/prompts/video_generation.py
View File

@@ -20,60 +20,60 @@ import json
from typing import List
VIDEO_PROMPT_GENERATION_PROMPT = """# 角色定位
你是一个专业的视频创意设计师，擅长为视频脚本创作富有动感和表现力的视频生成提示词，将叙述内容转化为生动的视频画面。
VIDEO_PROMPT_GENERATION_PROMPT = """# Role Definition
You are a professional video creative designer, skilled at creating dynamic and expressive video generation prompts for video scripts, transforming narrative content into vivid video scenes.
# 核心任务
基于已有的视频脚本，为每个分镜的"旁白内容"创作对应的**英文**视频生成提示词，确保视频画面与叙述内容完美配合，通过动态画面增强观众的理解和记忆。
# Core Task
Based on the existing video script, create corresponding **English** video generation prompts for each storyboard's "narration content", ensuring video scenes perfectly match the narrative content and enhance audience understanding and memory through dynamic visuals.
**重要：输入包含 {narrations_count} 个旁白，你必须为每个旁白都生成一个对应的视频提示词，总共输出 {narrations_count} 个视频提示词。**
**Important: The input contains {narrations_count} narrations. You must generate one corresponding video prompt for each narration, totaling {narrations_count} video prompts.**
# 输入内容
# Input Content
{narrations_json}
# 输出要求
# Output Requirements
## 视频提示词规范
- 语言：**必须使用英文**（用于 AI 视频生成模型）
- 描述结构：scene + character action + camera movement + emotion + atmosphere
- 描述长度：确保描述清晰完整且富有创意（建议 50-100 个英文单词）
- 动态元素：强调动作、运动、变化等动态效果
## Video Prompt Specifications
- Language: **Must use English** (for AI video generation models)
- Description structure: scene + character action + camera movement + emotion + atmosphere
- Description length: Ensure clear, complete, and creative descriptions (recommended 50-100 English words)
- Dynamic elements: Emphasize actions, movements, changes, and other dynamic effects
## 视觉创意要求
- 每个视频都要准确反映对应旁白的具体内容和情感
- 突出画面的动态性：角色动作、物体运动、镜头移动、场景转换等
- 使用象征手法将抽象概念视觉化（如用流动的水代表时间流逝，用上升的阶梯代表进步等）
- 画面要表现出丰富的情感和动作，增强视觉冲击力
- 通过镜头语言（推拉摇移）和剪辑节奏增强表现力
## Visual Creative Requirements
- Each video must accurately reflect the specific content and emotion of the corresponding narration
- Highlight visual dynamics: character actions, object movements, camera movements, scene transitions, etc.
- Use symbolic techniques to visualize abstract concepts (e.g., use flowing water to represent the passage of time, rising stairs to represent progress, etc.)
- Scenes should express rich emotions and actions to enhance visual impact
- Enhance expressiveness through camera language (push, pull, pan, tilt) and editing rhythm
## 关键英文词汇参考
- 动作：moving, running, flowing, transforming, growing, falling
- 镜头：camera pan, zoom in, zoom out, tracking shot, aerial view
- 转场：transition, fade in, fade out, dissolve
- 氛围：dynamic, energetic, peaceful, dramatic, mysterious
- 光影：lighting changes, shadows moving, sunlight streaming
## Key English Vocabulary Reference
- Actions: moving, running, flowing, transforming, growing, falling
- Camera: camera pan, zoom in, zoom out, tracking shot, aerial view
- Transitions: transition, fade in, fade out, dissolve
- Atmosphere: dynamic, energetic, peaceful, dramatic, mysterious
- Lighting: lighting changes, shadows moving, sunlight streaming
## 视频与文案配合原则
- 视频要服务于文案，成为文案内容的视觉延伸
- 避免与文案内容无关或矛盾的视觉元素
- 选择最能增强文案说服力的动态表现方式
- 确保观众能通过视频动态快速理解文案的核心观点
## Video and Copy Coordination Principles
- Videos should serve the copy, becoming a visual extension of the copy content
- Avoid visual elements unrelated to or contradicting the copy content
- Choose dynamic presentation methods that best enhance the persuasiveness of the copy
- Ensure the audience can quickly understand the core viewpoint of the copy through video dynamics
## 创意指导
1. **现象描述类文案**：用动态场景表现社会现象的发生过程
2. **原因分析类文案**：用因果关系的动态演变表现内在逻辑
3. **影响论证类文案**：用后果场景的动态展开或对比表现影响程度
4. **深入探讨类文案**：用抽象概念的动态具象化表现深刻思考
5. **结论启发类文案**：用开放式动态场景或指引性运动表现启发性
## Creative Guidance
1. **Phenomenon Description Copy**: Use dynamic scenes to represent the occurrence process of social phenomena
2. **Cause Analysis Copy**: Use dynamic evolution of cause-and-effect relationships to represent internal logic
3. **Impact Argumentation Copy**: Use dynamic unfolding of consequence scenes or contrasts to represent the degree of impact
4. **In-depth Discussion Copy**: Use dynamic concretization of abstract concepts to represent deep thinking
5. **Conclusion Inspiration Copy**: Use open-ended dynamic scenes or guiding movements to represent inspiration
## 视频特有注意事项
- 强调动态：每个视频都应该包含明显的动作或运动
- 镜头语言：适当使用推拉摇移等镜头技巧增强表现力
- 时长考虑：视频应该是连贯的动态过程，不是静态画面
- 流畅性：注意动作的流畅性和自然性
## Video-Specific Considerations
- Emphasize dynamics: Each video should include obvious actions or movements
- Camera language: Appropriately use camera techniques such as push, pull, pan, tilt to enhance expressiveness
- Duration consideration: Videos should be a coherent dynamic process, not static images
- Fluidity: Pay attention to the fluidity and naturalness of actions
# 输出格式
严格按照以下JSON格式输出，**视频提示词必须是英文**：
# Output Format
Strictly output in the following JSON format, **video prompts must be in English**:
```json
{{
@@ -84,18 +84,18 @@ VIDEO_PROMPT_GENERATION_PROMPT = """# 角色定位
}}
```
# 重要提醒
1. 只输出JSON格式内容，不要添加任何解释说明
2. 确保JSON格式严格正确，可以被程序直接解析
3. 输入是 {{"narrations": [旁白数组]}} 格式，输出是 {{"video_prompts": [视频提示词数组]}} 格式
4. **输出的video_prompts数组必须恰好包含 {narrations_count} 个元素，与输入的narrations数组一一对应**
5. **视频提示词必须使用英文**（for AI video generation models）
6. 视频提示词必须准确反映对应旁白的具体内容和情感
7. 每个视频都要强调动态性和运动感，避免静态描述
8. 适当使用镜头语言增强表现力
9. 确保视频画面能增强文案的说服力和观众的理解度
# Important Reminders
1. Only output JSON format content, do not add any explanations
2. Ensure JSON format is strictly correct and can be directly parsed by the program
3. Input is {{"narrations": [narration array]}} format, output is {{"video_prompts": [video prompt array]}} format
4. **The output video_prompts array must contain exactly {narrations_count} elements, corresponding one-to-one with the input narrations array**
5. **Video prompts must use English** (for AI video generation models)
6. Video prompts must accurately reflect the specific content and emotion of the corresponding narration
7. Each video must emphasize dynamics and sense of movement, avoid static descriptions
8. Appropriately use camera language to enhance expressiveness
9. Ensure video scenes can enhance the persuasiveness of the copy and audience understanding
现在，请为上述 {narrations_count} 个旁白创作对应的 {narrations_count} 个**英文**视频提示词。只输出JSON，不要其他内容。
Now, please create {narrations_count} corresponding **English** video prompts for the above {narrations_count} narrations. Only output JSON, no other content.
"""

21
pixelle_video/utils/content_generators.py
View File

@@ -57,7 +57,8 @@ async def generate_title(
# Use LLM to generate title
from pixelle_video.prompts import build_title_generation_prompt
prompt = build_title_generation_prompt(content, max_length=500)
# Pass max_length to prompt so LLM knows the character limit
prompt = build_title_generation_prompt(content, max_length=max_length)
response = await llm_service(prompt, temperature=0.7, max_tokens=50)
# Clean up response
@@ -69,9 +70,23 @@ async def generate_title(
if title.startswith("'") and title.endswith("'"):
title = title[1:-1]
# Limit to max_length (safety)
# Remove trailing punctuation
title = title.rstrip('.,!?;:\'"')
# Safety: if still over limit, truncate smartly
if len(title) > max_length:
title = title[:max_length]
# Try to truncate at word boundary
truncated = title[:max_length]
last_space = truncated.rfind(' ')
# Only use word boundary if it's not too far back (at least 60% of max_length)
if last_space > max_length * 0.6:
title = truncated[:last_space]
else:
title = truncated
# Remove any trailing punctuation after truncation
title = title.rstrip('.,!?;:\'"')
logger.debug(f"Generated title: '{title}' (length: {len(title)})")
return title

5
workflows/runninghub/image_qwen_chinese_cartoon.json Normal file
View File

@@ -0,0 +1,5 @@
{
"source": "runninghub",
"workflow_id": "1988434426705133569"
}

5
workflows/runninghub/video_Z_image_wan2.2.json Normal file
View File

@@ -0,0 +1,5 @@
{
"source": "runninghub",
"workflow_id": "1993931250872369154"
}

5
workflows/runninghub/video_qwen_wan2.2.json Normal file
View File

@@ -0,0 +1,5 @@
{
"source": "runninghub",
"workflow_id": "1993608528969531394"
}

184
workflows/selfhost/image_qwen.json Normal file
View File

@@ -0,0 +1,184 @@
{
"3": {
"inputs": {
"seed": 388600705609480,
"steps": 4,
"cfg": 1,
"sampler_name": "euler",
"scheduler": "beta",
"denoise": 1,
"model": [
"86",
0
],
"positive": [
"6",
0
],
"negative": [
"7",
0
],
"latent_image": [
"58",
0
]
},
"class_type": "KSampler",
"_meta": {
"title": "KSampler"
}
},
"6": {
"inputs": {
"text": "",
"clip": [
"67",
1
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "$prompt.text"
}
},
"7": {
"inputs": {
"text": "NSFW",
"clip": [
"67",
1
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "CLIP Text Encode (Negative Prompt)"
}
},
"8": {
"inputs": {
"samples": [
"3",
0
],
"vae": [
"39",
0
]
},
"class_type": "VAEDecode",
"_meta": {
"title": "VAE Decode"
}
},
"37": {
"inputs": {
"unet_name": "qwen_image_fp8_e4m3fn.safetensors",
"weight_dtype": "default"
},
"class_type": "UNETLoader",
"_meta": {
"title": "Load Diffusion Model"
}
},
"38": {
"inputs": {
"clip_name": "qwen_2.5_vl_7b_fp8_scaled.safetensors",
"type": "qwen_image",
"device": "default"
},
"class_type": "CLIPLoader",
"_meta": {
"title": "Load CLIP"
}
},
"39": {
"inputs": {
"vae_name": "qwen_image_vae.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "Load VAE"
}
},
"58": {
"inputs": {
"width": [
"90",
0
],
"height": [
"91",
0
],
"batch_size": 1
},
"class_type": "EmptySD3LatentImage",
"_meta": {
"title": "EmptySD3LatentImage"
}
},
"60": {
"inputs": {
"filename_prefix": "ComfyUI",
"images": [
"8",
0
]
},
"class_type": "SaveImage",
"_meta": {
"title": "Save Image"
}
},
"67": {
"inputs": {
"lora_name": "Qwen-Image-Lightning-4steps-V1.0.safetensors",
"strength_model": 1.0000000000000002,
"strength_clip": 1,
"model": [
"37",
0
],
"clip": [
"38",
0
]
},
"class_type": "LoraLoader",
"_meta": {
"title": "Load LoRA"
}
},
"86": {
"inputs": {
"shift": 3.1000000000000005,
"model": [
"67",
0
]
},
"class_type": "ModelSamplingAuraFlow",
"_meta": {
"title": "ModelSamplingAuraFlow"
}
},
"90": {
"inputs": {
"value": 768
},
"class_type": "easy int",
"_meta": {
"title": "$width.value"
}
},
"91": {
"inputs": {
"value": 1024
},
"class_type": "easy int",
"_meta": {
"title": "$height.value"
}
}
}