搜索 K
Appearance
Appearance
手工逐条写Prompt是作坊式生产,利用“LLM动态变量注入 + 状态机自动化管线”才是工业化未来:采用DeepSeek-R1 / V3作为结构化编译器,注入角色特征全局常量(锚定服装、发型、面容UUID)、场景时空状态机(昼夜、天气、战损度)与机位运动参数,自动输出标准JSON分镜Prompt流。结合Python脚本与ComfyUI API,单部80集短剧可在2小时内完成上千条镜头Prompt的编译。底层批量渲染挂载于光速云GPU算力容器(邀请码 keqgvT5Y),成片快速对接黄果短剧官方商用生态实现全自动化量产变现。
对于制作80到100集连续短剧的工作室而言,纯人工编写分镜提示词面临着三大致命缺陷:耗费海量人力、角色服装道具在前后集数中频繁漂移、风格描述前后矛盾。
本篇面向技术型导演与工作室技术负责人,详细拆解一套基于大语言模型动态注入与多轮上下文状态机的高阶短剧Prompt工程架构。
flowchart TD
Novel[3万字原创短剧剧本] --> StateMachine[全局状态机管理器 Global State Machine]
subgraph 全局状态存储器
C_DB[(角色库: UUID/服装/身材/声线)]
S_DB[(场景库: 豪宅/雨夜/破庙/会议室)]
T_DB[(时间线状态: 战损值/情绪值/道具持有)]
end
Novel --> LLM[DeepSeek-R1 结构化推理引擎]
C_DB -.-> LLM
S_DB -.-> LLM
T_DB -.-> LLM
LLM --> JSONStream[标准分镜JSON流: Shot-by-Shot Prompts]
JSONStream --> APIEngine[ComfyUI / 视频模型批量API]
APIEngine --> CloudCompute[光速云算力集群并发渲染]
CloudCompute --> MasterDelivery[黄果短剧变现网络 huangguoju.co]character_manifest.json) {
"characters": {
"protagonist_lu": {
"name": "陆景深",
"base_prompt": "30-year-old East Asian male, chiselled jawline, short fade black hair, intense stoic amber eyes",
"costume_default": "tailored bespoke charcoal three-piece suit, crisp white dress shirt, platinum tie clip",
"costume_battle": "shredded tactical combat trench coat, carbon-fiber armor chest plate, blood stains on shoulder",
"negative": "beard, stubble, soft jaw, round eyes, cartoonish"
},
"heroine_su": {
"name": "苏清浅",
"base_prompt": "24-year-old delicate East Asian female, porcelain skin, long wavy ebony hair, red tearful eyes",
"costume_default": "emerald green silk evening gown, diamond droplet earrings, silver heels",
"negative": "blonde hair, heavy western features, heavy dark makeup"
}
}
}向DeepSeek-R1注入工业系统级Prompt,使其按标准Schema输出每个镜头的全套参数:
你是一位好莱坞级影视视觉总监与AI短剧分镜架构师。
请严格读取输入的剧情片段与角色状态清单,输出符合以下JSON Schema的分镜Prompt:
{
"shot_id": "EP01_S01",
"character_ref": "protagonist_lu",
"state_condition": "battle_damaged",
"camera": {
"angle": "Low-angle tilted 15 degrees",
"lens": "35mm anamorphic prime",
"movement": "Smooth dolly-in"
},
"lighting": "High-contrast rim lighting, cold blue street lamps with volumetric rain mist",
"action_prompt_en": "Cinematic vertical 9:16, [protagonist_lu.base_prompt], wearing [protagonist_lu.costume_battle], slowly wiping blood from split lip with back of hand, glaring with predatory coldness toward off-screen camera. Background of torrential rain at luxurious estate gate.",
"negative_prompt": "smiling, cross-eyed, deformed fingers, cartoon"
}# 状态机自动注入与编译脚本
import json
def compile_shot_prompt(manifest_file, raw_shot_plan):
with open(manifest_file, 'r', encoding='utf-8') as f:
manifest = json.load(f)
compiled_shots = []
for shot in raw_shot_plan:
char_key = shot["character_ref"]
char_info = manifest["characters"][char_key]
# 动态替换变量
pos_prompt = shot["action_prompt_en"]
pos_prompt = pos_prompt.replace("[protagonist_lu.base_prompt]", char_info["base_prompt"])
costume = char_info["costume_battle"] if shot["state_condition"] == "battle_damaged" else char_info["costume_default"]
pos_prompt = pos_prompt.replace("[protagonist_lu.costume_battle]", costume)
compiled_shots.append({
"shot_id": shot["shot_id"],
"positive_prompt": pos_prompt,
"negative_prompt": char_info["negative"] + ", " + shot["negative_prompt"],
"camera": shot["camera"]
})
return compiled_shots| 流程指标 | 传统纯人工手工写Prompt | 状态机动态变量编译方案 | 光速云并发自动化方案 |
|---|---|---|---|
| 80集全剧Prompt编写耗时 | 5 - 7 天 (编剧极易疲劳失误) | 2 - 3 小时 (代码秒级生成) | 30分钟批量完成 |
| 角色服装发型一致性 | 仅 45% (频繁穿帮变形) | 95% 以上 (强变量锁定) | 98% (配合LoRA与Seed锁定) |
| 机位与景深合规度 | 参差不齐,容易遗漏镜头语言 | 100% 工业级标准化约束 | 100% 批量化质检过滤 |
| 全剧算力损耗成本 | 抽卡废品率高达 50% | 废品率压至 15% 以内 | 单集成本低至数元 |