Tokyo Rain Chase — Context-IR Side-by-Side

MiniMax H3CarActionWorkflowRain Scene15s
The same 15-second rainy-night car chase prompt, compared side by side: the original 2K output versus the Context-IR-rewritten, super-resolved version. 20 shot changes plus a slow-motion water splash — a real-world test sample for a prompt-rewriting tool.
PROMPT · 原始视频提示词
Create with AI
The original prompt for this 15-second car chase video is as follows: This is a Hollywood action movie clip, an urban car chase film. The time is after rain, at night, and the location is the streets of Tokyo. Two cars chase each other through the streets, and finally, on a bridge in Tokyo, the orange sports car behind rams the black Mercedes in front off the bridge. The black car falls into the water, and the orange car successfully speeds away. Within the 15 seconds there should be 20 shot/angle cuts, and slow motion should be used when the car falls into the water. After using Context-IR, a new prompt was generated, which I've posted in the comments. For the original prompt I generated 2K directly; for the Context-IR version I first generated it at 768P and then did official super-resolution upscaling. I did a side-by-side comparison of the two videos — which one do you think is better? 0:08 / 0:
✍️ Editor’s Notes

这条其实不是标准视频提示词,而是一条原始提示词对比Context-IR处理效果的讨论帖,本体信息量很薄,只给了一句概括——东京雨夜追车、撞车坠河、二十个景别切换,真正的关键词密度和结构细节被作者甩进了评论区,单独拎出来看是不完整的。放在MiniMax H3上做对比也说得通,这类模型对提示词精炼度敏感,作者想验证的是人工精修的短提示词能不能打过工具预处理再超分的路径,属于工作流实验记录而非纯创作提示词。

Want tighter control? Editor-expanded reference version (not the original prompt):
好莱坞动作片质感的都市夜间飙车片段,时长15秒。雨后东京街头,路面反光如镜,霓虹灯在积水中碎成光斑。黑色奔驰在前,橙色跑车在后,高速追逐中轮胎甩起水雾,二十个不同景别镜头快速切换——低角度贴地跟拍、车内反打、俯拍航拍轮番出现,剪辑干脆不拖沓。高潮落在东京跨江大桥,橙色跑车侧撞逼停黑色奔驰,奔驰被撞出护栏坠入江中,此刻切入慢动作,水花四溅、车身翻滚、玻璃碎裂逐帧可见,声音短暂抽离制造悬浮感。橙色跑车不作停留,加速驶离桥面,尾灯没入雨夜街道尽头,画面恢复正常速率收尾。
Create with AI
Related Works
×