TemplateCraft: Agentic Visual Template Generation

Hongjie Yu; Zhiyuan Fan; Yuzhe Zhang; Jiangcun Du; Zhicheng Gao; Yuhong Zhang; Xiaokai Zhan; Zongshi Xie

cs.MMSep 25, 2026

TemplateCraft: Agentic Visual Template Generation

Authors: Hongjie Yu, Zhiyuan Fan, Yuzhe Zhang, Jiangcun Du, Zhicheng Gao, Yuhong Zhang, Xiaokai Zhan, Zongshi Xie

Organizations: Peking University · Kuaishou Technology

Abstract

The growing popularity of short videos has driven demand for one-click content creation. Visual templates turn uploaded images into personalized content with preset effects, but reusable template generation still requires substantial manual effort in asset preparation and tool orchestration. We propose TemplateCraft, a multi-agent system that converts natural-language instructions into client-executable templates through planning, material generation, effect-workflow generation, and protocol compilation. Its Planner-Evaluator loop uses execution feedback for targeted rollback, while stage-level and long-term memory support revision without parameter updates. We evaluate TemplateCraft on TemplateBench, derived from 60 real-world templates. With the same Qwen3-VL backbone, TemplateCraft raises image/video generation success rates from 56.7%/30.0% to 66.7%/50.0% over Planner-only (best-of-three) and improves template adherence and style consistency. With additional evaluation and revision, it matches or exceeds a GPT-4o Planner-only baseline on selected metrics. Persistent assets further improve cross-input style consistency.

Figures & tables

Figure 1: The TemplateCraft multi-agent loop for template generation and template reuse.

Method	Video Generation	Template Generation	Image Template Tasks					Video Template Tasks
Method	Video Generation	Template Generation	$\mathrm{SR}_{c}$	TA	TR	IQ	AQ	$\mathrm{SR}_{c}$	TA	TR	IQ	AQ
Wan2.6	$\checkmark$		–	0.8275	0.6411	0.7301	0.6009	–	0.4283	0.7185	0.7344	0.6263
Seed	$\checkmark$		–	0.8467	0.6417	0.7289	0.6018	–	0.5118	0.6651	0.7201	0.6050
Qwen3 + UniVA	$\checkmark$		–	0.7733	0.6226	0.7391	0.5914	–	0.7164	0.6966	0.7184	0.6554
Qwen3 + Planner-only	$\checkmark$	$\checkmark$	56.7%	0.5168	0.6957	0.7186	0.5693	30.0%	0.4034	0.7526	0.7105	0.6574
Qwen3 + Planner-CoT	$\checkmark$	$\checkmark$	56.7%	0.5675	0.7151	0.7345	0.5636	43.3%	0.5481	0.7913	0.7109	0.6544

Table 1: Evaluation results of all methods on TemplateBench. Higher is better for all metrics; checkmarks indicate supported generation capabilities, and – denotes a metric that is not applicable. Bold and underlined values indicate the best and second-best results, respectively.

Figure 2: An evaluation task with a creative brief, five test-input sets, and a scoring rubric.

Figure 3: TemplateCraft ablation and capability comparison, together with the persistent-asset experiment. Left: template adherence (TA) and template generation success rate (

\mathrm{SR}_{c}

) for different system configurations and the GPT-4o baseline on image and video tasks. Right: template reusability (TR) with and without an explicit requirement to use persistent assets.

Tool-Using Vision-Language Agents

Tool-using vision-language agents combine visual and language models with external tools to gather evidence, act, and solve multi-step tasks.

See all 140 papers

FRAMEWORKERS: A Dynamic Multi-Agent Framework for AI-Generated Video Production
Aug 30, 2026Zhendong Li, Lei Sun, Letian Shi +8Multi-Agent Orchestration Agentic Video Generation
INSAIT, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria · Snap Inc., New York, USA
VisionCreator: A Native Visual-Generation Agentic Model with Understanding, Thinking, Planning and Creation
Mar 3, 2026Jinxiang Lai, Zexin Lu, Jiajun He +11Agentic Image Generation
Hong Kong University of Science and Technology · Tencent Hunyuan
Open Evaluation Agent: Efficient and Promptable Evaluation of Visual Generative Models
Aug 10, 2026Shulin Tian, Ziqi Huang, Fan Zhang +3Tool-Augmented Language Model Agents AI Agent Evaluation
Shanghai Artificial Intelligence Laboratory

Aug 30, 2026J/K move · Enter open · S save

FRAMEWORKERS: A Dynamic Multi-Agent Framework for AI-Generated Video Production

Zhendong Li, Lei Sun, Letian Shi, Deheng Zhang+7

INSAIT, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria · Snap Inc., New York, USA

Multi-Agent Orchestration Agentic Video Generation LLM Agent Orchestration Video Generation

Open paper AskPDF arXiv

Modern video generators excel at synthesizing individual clips, but complete video production requires coordinating a long sequence of interdependent creative steps, including scripting, storyboarding, generation, and editing. It further demands persistent asset management and dynamic task orchestration as intermediate outputs, dependencies, and execution states evolve over time. Existing automated systems typically rely on rigid pipelines that are difficult to adapt to diverse inputs and changing workflows, while general-purpose large language models (LLMs) remain unreliable for long-horizon orchestration and multimodal asset routing. We introduce FRAMEWORKERS, a task-centric and workspace-grounded multi-agent framework for open-ended video production. A central Director formulates video creation as dynamic task management, continuously editing a Task Stack to determine which subtask to execute next and which sub-agent to invoke. An Assistant serves as the execution layer, grounding each selected task in a shared Workspace, retrieving the required assets and context, invoking the assigned sub-agent, and persisting the resulting artifacts. Execution capabilities are exposed through modular sub-agents with registered descriptors, allowing new sub-agents to be integrated without redesigning the orchestration workflow. To improve orchestration reliability, we fine-tune the Director via supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) for descriptor-conditioned task routing. Experiments show that FRAMEWORKERS outperforms strong LLM planners in routing accuracy, recovers reliably from runtime failures, generalizes to unseen sub-agents without retraining, and achieves higher end-to-end video quality and broader task coverage than fixed pipelines, single-agent systems, and prior multi-agent approaches.

Abstract

Tool-Using Vision-Language Agents

Explore similar work

FRAMEWORKERS: A Dynamic Multi-Agent Framework for AI-Generated Video Production

VisionCreator: A Native Visual-Generation Agentic Model with Understanding, Thinking, Planning and Creation

Open Evaluation Agent: Efficient and Promptable Evaluation of Visual Generative Models