README.md

metadata

base_model: XD-MU/ScriptAgent
library_name: peft
pipeline_tag: text-generation
tags:
  - base_model:adapter:XD-MU/ScriptAgent
  - lora
  - transformers
arxiv: 2601.17737

ScriptAgent: Dialogue-to-Shooting-Script Generation Model

This model is a fine-tuned adapter (LoRA) designed to generate detailed shooting scripts from dialogue inputs. It is the implementation of ScripterAgent as described in the paper: The Script is All You Need: An Agentic Framework for Long-Horizon Dialogue-to-Cinematic Video Generation.

Project Page | Code | Demo

Model Description

ScriptAgent transforms conversational text (coarse dialogue) into structured, fine-grained, and executable cinematic scripts. It bridges the "semantic gap" between a creative idea and its cinematic execution, providing necessary context for video generation models, including character descriptions, scene settings, positions, and dialogue cues.

The model is compatible with ms-swift and supports efficient inference via the vLLM backend.

💡 Note: This repository contains a PEFT adapter (LoRA). To use it, you must merge it with the original base model or load it via ms-swift.

▶️ Inference with ms-swift (vLLM Backend)

To generate shooting scripts from dialogue inputs, use the following snippet with ms-swift. You can find DialoguePrompts here.

import os
from huggingface_hub import snapshot_download
from swift.llm import PtEngine, RequestConfig, InferRequest

os.environ['CUDA_VISIBLE_DEVICES'] = '0'

model_name = "XD-MU/ScriptAgent"
local_path = "./models/ScriptAgent"

# Download the model files
print("Downloading model...")
snapshot_download(
    repo_id=model_name,
    local_dir=local_path,
    local_dir_use_symlinks=False,
    resume_download=True
)

# Load using SWIFT
engine = PtEngine(local_path, max_batch_size=1)
request_config = RequestConfig(max_tokens=8192, temperature=0.7)

infer_request = InferRequest(messages=[
    {"role": "user", "content": "Your Dialogue Here"}
])
response = engine.infer([infer_request], request_config)[0]

print(response.choices[0].message.content)

Citation

If you find this work useful, please cite:

@article{directing2026,
  title={The Script is All You Need: An Agentic Framework for Long-Horizon Dialogue-to-Cinematic Video Generation},
  author={Mu, Chenyu and He, Xin and Yang, Qu and Chen, Wanshun and Yao, Jiadi and Liu, Huang and Yi, Zihao and Zhao, Bo and Chen, Xingyu and Ma, Ruotian and others},
  journal={arXiv preprint arXiv:2601.17737},
  year={2026}
}

Acknowledgments

• Thanks to LLaMA-Factory for the SFT training framework.
• Thanks to ms-swift for the GRPO training framework.