image-to-video

安装 belt CLI 技能： npx skills add belt-sh/cli

图片转视频

通过 inference.sh CLI 将静态图片转换为动画视频。

快速开始

需要 inference.sh CLI（belt）。安装说明

belt login

# Generate a still image
belt app run falai/flux-dev-lora --input '{
  "prompt": "serene mountain lake at sunset, snow-capped peaks reflected in still water, golden hour light, landscape photography",
  "width": 1248,
  "height": 832
}'

# Animate it
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "gentle ripples on the lake surface, clouds slowly drifting, warm light shifting, birds flying in the distance",
  "image": "path/to/lake-image.png"
}'

模型选择

模型	应用 ID	最适合	运动风格
Wan 2.5 i2v	falai/wan-2-5-i2v	写实运动、自然动作	照片级写实、细腻
WAN-I2V (Pruna)	pruna/wan-i2v	经济、快速、480p/720p	自然、高效
Seedance 2.0	bytedance/seedance-2-0	最高支持 1080p、音画同步、支持所有输入类型	多样、高质量
Seedance 2.0 Fast	bytedance/seedance-2-0-fast	快速版本，功能相同	多样、快速
Fabric 1.0	falai/fabric-1-0	布料、织物、液体、流动材质	基于物理的流动效果
Grok Imagine Video	xai/grok-imagine-video	通用动画、文本引导	多样

何时使用各模型

场景	最佳模型	原因
有水或云的风景	Wan 2.5 i2v	最擅长自然、逼真的运动
表情细微的人像	Wan 2.5 i2v	保持面部特征一致
带有布料或织物的产品	Fabric 1.0	专精于材质物理效果
飘扬的旗帜、流动的窗帘	Fabric 1.0	布料模拟
插画或艺术风格图片	Seedance 2.0	与风格化内容相匹配
通用的“让画面活起来”	Seedance 2.0	综合表现出色，最高支持 1080p
快速测试或迭代	Seedance 2.0 Fast	生成速度更快

运动类型

镜头运动

运动	提示词关键词	效果
推近 / 向前移动	"slow dolly forward", "camera pushes in"	增强亲近感或聚焦效果
拉远 / 向后移动	"camera pulls back", "slow zoom out"	展现更多画面和背景
向左或向右摇摄	"camera pans slowly to the right"	扫视、跟随
向上或向下倾斜	"camera tilts upward"	展现高度
环绕	"camera orbits around the subject"	3D 探索
升降镜头上升	"camera rises upward"	展现宏大场景
固定	（不使用镜头运动提示词）	仅主体运动

主体运动

类型	提示词示例
自然元素	"water rippling", "clouds drifting", "leaves rustling in breeze"
头发或衣物	"hair blowing gently in wind", "dress fabric flowing"
氛围效果	"fog slowly rolling", "dust particles floating in light beams"
人物	"person slowly turns to camera", "subtle breathing motion"
机械	"gears turning", "clock hands moving"
液体	"coffee steam rising", "paint dripping", "water pouring"

提示词最佳实践

黄金法则：细腻胜过夸张

AI 视频模型对轻柔、细腻的运动的生成效果优于夸张的动作。要求过多运动会导致变形和瑕疵。

❌ "person running and jumping over obstacles while the camera spins"
✅ "person slowly walking forward, gentle breeze, camera follows alongside"

❌ "explosion with debris flying everywhere"
✅ "candle flame flickering gently, warm ambient light shifting"

❌ "fast zoom into the eyes with dramatic camera shake"
✅ "slow dolly forward toward the subject, subtle focus shift"

提示词结构

[Camera movement] + [Subject motion] + [Atmospheric effects] + [Mood/pace]

按场景分类的示例

# Landscape animation
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "gentle camera pan right, water reflecting moving clouds, trees swaying slightly in breeze, warm golden light, peaceful and slow",
  "image": "landscape.png"
}'

# Portrait animation
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "subtle breathing motion, slight head turn, natural eye blink, hair moving gently, soft ambient lighting shifts",
  "image": "portrait.png"
}'

# Product shot animation
belt app run bytedance/seedance-2-0 --input '{
  "prompt": "slow 360 degree orbit around the product, gentle spotlight movement, subtle reflections shifting, premium product showcase, smooth motion",
  "image": "product.png",
  "generate_audio": true
}'

# Fabric/cloth animation
belt app run falai/fabric-1-0 --input '{
  "prompt": "fabric flowing and rippling in gentle wind, natural cloth physics, soft movement",
  "image": "fabric-scene.png"
}'

# Architectural visualization
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "slow dolly forward through the entrance, slight camera tilt upward, ambient light filtering through windows, dust particles in light beams",
  "image": "building-interior.png"
}'

时长指南

时长	质量	用途
2-3 秒	最高质量	GIF、循环背景、动态照片
4-5 秒	高质量	社交媒体帖子、产品展示
6-8 秒	良好质量	短片、转场
10 秒以上	质量下降	除非拼接较短片段，否则避免使用

延长时长

要生成更长的视频，可生成多个较短片段并进行拼接：

# Generate 3 clips from the same image with progressive motion
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "slow pan left, gentle water motion",
  "image": "scene.png"
}' --no-wait

belt app run falai/wan-2-5-i2v --input '{
  "prompt": "continuing pan, clouds shifting, light changing",
  "image": "scene.png"
}' --no-wait

# Stitch together
belt app run infsh/media-merger --input '{
  "media": ["clip1.mp4", "clip2.mp4"]
}'

完整工作流

从静态图到最终视频的流程

# 1. Generate source image (best quality)
belt app run bytedance/seedream-4-5 --input '{
  "prompt": "cinematic landscape, misty mountains at dawn, lake in foreground, dramatic clouds, golden hour, 4K quality, professional photography",
  "size": "2K"
}'

# 2. Animate the image
belt app run falai/wan-2-5-i2v --input '{
  "prompt": "gentle mist rolling through the valley, lake surface rippling, clouds slowly moving, birds in distance, warm light shifting",
  "image": "landscape.png"
}'

# 3. Upscale video if needed
belt app run falai/topaz-video-upscaler --input '{
  "video": "animated-landscape.mp4"
}'

# 4. Add ambient audio
belt app run infsh/hunyuanvideo-foley --input '{
  "video": "animated-landscape.mp4",
  "prompt": "gentle nature ambience, distant birds, soft wind, water lapping"
}'

# 5. Merge video with audio
belt app run infsh/video-audio-merger --input '{
  "video": "upscaled-landscape.mp4",
  "audio": "ambient-audio.mp3"
}'

Cinemagraph 效果

Cinemagraph 是一种静态照片，其中只有一个元素在移动（例如，瀑布在静止的场景中流动）。要实现这种效果：

1. 生成静态图像，并清晰呈现运动元素
2. 仅提示该特定元素运动
3. 时长保持在 2-4 秒，以实现无缝循环

belt app run falai/wan-2-5-i2v --input '{
  "prompt": "only the waterfall is moving, everything else remains perfectly still, water cascading smoothly, rest of scene frozen",
  "image": "waterfall-scene.png"
}'

常见错误

错误	问题	修正方法
要求过多运动	失真、伪影、扭曲	始终选择微妙而非夸张的运动
内容类型与模型不匹配	效果不佳	使用上方的选择指南
片段过长（10 秒以上）	质量会显著下降	时长保持在 3-5 秒，需要时拼接片段
未指定摄像机运动	运动随机且不可预测	始终指定摄像机行为
运动方向相互冲突	混乱、不自然	只指定一个主要运动方向
源图像分辨率低	视频输出分辨率低	从质量最高的源图像开始
动作场景过于复杂	模型无法处理	让运动保持简单、自然

相关技能

npx skills add inference-sh/skills@ai-video-generation
npx skills add inference-sh/skills@ai-image-generation
npx skills add inference-sh/skills@p-video
npx skills add inference-sh/skills@video-prompting-guide
npx skills add inference-sh/skills@prompt-engineering

浏览所有应用：belt app list