XML-Tagged Prompting — A Sora 2 Experiment

Sora 2Structured PromptCreativeHandheld CameraFilm Grain10s
The prompt is written with XML tags like <video model="sora-2" duration="8s">, with look and style each broken into their own nodes. A third form of structured prompting — beyond JSON and YAML — rounding out the format collection.
PROMPT · Video Prompt
Create with AI
<video version="2.0" model="sora-2" duration="8s" fps="24" aspect_ratio="9:16" resolution="1080x1920" seed="auto"> <look style="cinematic realism" color_grade="muted greens and grays, creamy whites, gentle rose skin" grain="subtle"/> <environment> <location>Back-alley stoop with green-painted steel stairs, cracked pavement, small strip of grass and stones</location> <time>Night</time> <weather>Dry pavement with slight humidity sheen</weather> <lighting> <key>On-camera flash pop look</key> <fill>Cool streetlamp spill from camera-left</fill> <background>Falls to shadow, retain highlight detail on white wings</background> </lighting> </environment> <camera sensor="smartphone-front feel" lens="26-28mm equiv" shutter="1/48s cinematic, mild motion blur" aperture="f/2.2 shallow DOF with soft halation on speculars" stabilization="handheld"> <movement type="single push-in" path="straight ~0.5m" micro_tilt="minimal to keep face centered" wobble_limit="handheld micro-bob ≤ 3% frame height"/> <rules>One primary motion only, no zooms, no extra moves, no cuts</rules> </camera> <cast> <human id="subject01" age="young adult" hair="long black, straight fringe" makeup="pale with soft blush" pose="seated curl on green metal steps, hands clasping shoe laces, looking down" mood="melancholic, delicate, ethereal but grounded, PG-13"> <wardrobe>Black blazer over light dress with tulle pooling on ground, chunky brown platform oxfords, large white feathered costume wings</wardrobe> <notes>No logos; wings are practical costume, no flight</notes> </human> </cast> <composition> <framing>Subject centered; wings form diagonals; stairs and pavement add gritty texture</framing> <depth_of_field>Face and front wing plane sharp, background softly falls off</depth_of_field> <atmospherics>Subtle film grain; slight haze around highlights to echo flash bloom</atmospherics> <palette primary="industrial green, asphalt gray, soft ivory" accents="rose blush, brown leather shoes"/> </composition> <audio> <ambience>Faint urban hum, distant traffic hush, night insects</ambience> <foley>Shoe lace tug, fabric rustle, slight step creak</foley> <dialogue>None</dialogue> <music>None</music> </audio> <timeline> <event t="0.0-2.5s"> <shot type="medium from overhead 3/4 angle">Subject sits curled, wings slightly splayed</shot> <physics>Loose hair shifts subtly; feathers quiver from faint breeze</physics> <camera_motion>Begin slow handheld push-in</camera_motion> </event> <event t="2.5-5.0s"> <action>She tightens a lace, exhales; chin tilts, eyes glance down; wings remain still</action> <physics>Tulle settles with micro ripples; cloth compresses at knees</physics> </event> <event t="5.0-8.0s"> <shot type="tight portrait">Camera eases in; she lifts gaze slightly toward lens then softly closes eyes; hold on quiet</shot> <physics>Minimal hand tremor; stable specular roll on feathers and shoes</physics> </event> </timeline> <constraints no_text_on_screen="true" no_watermarks="true" tone="PG-13 respectful" vfx="no supernatural, wings are practical"/> <negative_prompt>cartoonish acting, rolling-shutter wobble, clipping on feathers or hair, overexposure, extra camera moves, zooms, cuts</negative_prompt> </video>
✍️ Editor’s Notes

用XML标签把镜头、光线、角色、时间轴一层层嵌套写清楚,这种写法在Sora 2的重度玩家里不算新鲜,本质是把自然语言提示词工程化,好处是每个参数都能单独调试,坏处是信息密度太高,模型未必会逐字执行,很多细分参数最后大概率被模型当成风格参考而不是精确指令。负面提示词那一栏写得挺全,卡通感、滚动快门抖动、镜头切换过多这些都是Sora 2实际生成里常见的翻车点,说明作者是踩过坑之后才总结出来的。天使少女蹲坐台阶这个画面本身氛围感不错,但整条提示词的复杂度已经超过了大部分场景真正需要的精细度。

Create with AI
Related Works
×