TemplateCraft is a multi-agent system that converts natural-language instructions into reusable visual templates for short videos. It decomposes the task into planning, material generation, effect-workflow generation, and protocol compilation stages, using a Planner-Evaluator loop with stage-level and long-term memory to guide revision without parameter updates.
Evaluated on TemplateBench (60 real-world templates from KwaiCut), TemplateCraft improves image/video generation success rates from 56.7%/30.0% (Planner-only, best-of-three) to 66.7%/50.0% using the same Qwen3-VL backbone. It also improves template adherence and style consistency, and with additional revision matches or exceeds a GPT-4o Planner-only baseline on selected metrics. Persistent assets further improve cross-input style consistency.
This work addresses the manual effort required in creating visual templates, a key bottleneck for personalized one-click content creation in the short-video domain.