搜索 K
Appearance
Appearance
主流大模型的底层文本编码器(Text Encoder)存在巨大代际差异,生搬硬套必炸模:Midjourney依赖短语权重与CLI参数标签(如 `--ar 9:16 --c 5 --sref`);Stable Diffusion/ComfyUI依赖CLIP权重符号 `(word:1.3)` 与注意力打散;而新一代视频大模型(可灵2.0、Runway Gen-3、海螺AI)则基于T5-XXL或多模态大语言模型(MLLM),必须使用“主谓宾完整自然句 + 物理动力学描述”。掌握跨模型转译编译规则,一套剧本可同时在绘图与视频模型中精准流转。批量跨模型调度时,推荐借助光速云GPU算力容器(邀请码 keqgvT5Y)秒级并发生成,最终成片直推黄果短剧官方分发矩阵实现全网变现。
在AI短剧工业流水线中,创作者通常会在不同环节使用不同工具:在策划阶段用Midjourney设计概念海报、在分镜原画阶段用ComfyUI(SDXL/FLUX)固定角色面容,最后在视频生成阶段使用可灵、海螺或Runway赋予动态。
如果把Midjourney里充满 --v 6 --ar 9:16 ::2 的咒语直接复制进可灵或Runway,视频模型将无法解析这些命令行参数,甚至会把参数识别为屏幕上的乱码字幕。
本文系统拆解三大主流提示词语法流派的底层机理与无损转译规则。
flowchart TD
Novel[剧本分镜语义源] --> Core[纯净视听描述核心 Canonical Audio-Visual Core]
Core --> Trans1[Midjourney 转译器: 提取艺术词 + 追加参数 --ar 9:16 --v 6.1 --s 250]
Core --> Trans2[SD/ComfyUI 转译器: 注入权重 (word:1.2) + LoRA 语法 + 负向词池]
Core --> Trans3[可灵/Runway 转译器: 扩展为多模态时空自然语言完整长句 + 运镜轨迹动词]
Trans1 --> Render1[Midjourney 概念海报]
Trans2 --> Render2[ComfyUI 批量定妆照]
Trans3 --> Render3[可灵/Runway 动态视频]
Render2 --> Cloud[光速云GPU批量渲染]
Render3 --> Cloud
Cloud --> Master[黄果短剧商业分发 huangguoju.co]| 体系流派 | 代表模型 | 核心编码器 (Text Encoder) | 最佳语法风格 | 绝不可用的语法 |
|---|---|---|---|---|
| 标签权重流 | SD 1.5, SDXL, ComfyUI | CLIP ViT-L / OpenCLIP | 逗号分隔的短语单词、明确圆括号权重 (word:1.2) | 复杂从句、长篇抒情段落 |
| 参数标记流 | Midjourney v6, Niji 6 | 专有魔改多模态模型 | 摄影短语 + 双冒号权重 word::2 + CLI参数 --ar | 传统代码块、中文乱码参数 |
| 时空自然叙事流 | 可灵AI, Runway Gen-3, 海螺AI | T5-XXL / 自研MLLM大模型 | 完整时态自然语言长句、包含物理因果与摄影机运镜 | 任何 -- 符号、数字权重括号、散乱词堆砌 |
目标镜头:暴雨夜,现代霸总男主手握婚书站在豪宅门前,冷漠推镜特写。
Cinematic 9:16 vertical, a 30-year-old Chinese billionaire CEO standing in heavy torrential rain outside an opulent mansion gate, holding a ripped antique marriage agreement, cold piercing gaze, soaked black tailored wool overcoat, blue car headlights rim light, reflections on wet asphalt, volumetric rain mist, 85mm portrait photography --ar 9:16 --v 6.1 --style raw --s 250(masterpiece:1.3), (best quality:1.2), (photorealistic:1.3), 1boy, 30-year-old Chinese male, (cold amber eyes:1.2), (sharp jawline:1.2), wearing (tailored charcoal wool overcoat:1.2), holding (torn paper contract:1.1), standing in (torrential rain:1.2) outside luxury European mansion gate, puddle reflections, (chiaroscuro lighting:1.2), (rim light:1.3), shot on 85mm lens, <lora:CEO_Male_Face_v2:0.8>
Negative prompt: (worst quality:1.4), (low quality:1.4), deformed hands, missing fingers, cross-eyed, plastic skin, cartoon, animeA cinematic slow dolly-in tracking shot moving smoothly toward the protagonist. A 30-year-old Chinese billionaire CEO in a rain-soaked black wool overcoat stands motionless in a heavy downpour outside a grand mansion gate. In his steady hand, he holds a torn marriage contract as rain drips from his fingertips. His cold, unblinking eyes glare directly toward the camera with suppressed hostility. Cold blue vehicle headlights illuminate the pouring rain from behind, creating sharp cinematic rim lighting on his shoulders. 8k, photorealistic 24fps.# 自动化将标准描述转为不同模型格式
def compile_prompt(scene_desc, target_platform="kling"):
if target_platform == "midjourney":
return f"{scene_desc['subject']}, {scene_desc['env']}, {scene_desc['lighting']} --ar 9:16 --v 6.1 --style raw"
elif target_platform == "sdxl":
return f"(masterpiece:1.2), {scene_desc['subject']}, in {scene_desc['env']}, ({scene_desc['lighting']}:1.2), 85mm lens, 8k"
elif target_platform == "kling" or target_platform == "runway":
return f"A cinematic {scene_desc['camera_motion']} shot focusing on {scene_desc['subject']} who is {scene_desc['action']} in {scene_desc['env']}. {scene_desc['lighting']}, hyper-realistic motion fluidity, 24fps."--ar 9:16 水印?--ar 9:16 当成了视觉文本描述,误认为你想在视频中画出这串字符。转译到视频模型前,必须全局正则表达式清除所有 -- 开头的参数。