Story2Board offers a training-free framework for generating expressive storyboards from natural language descriptions. It addresses limitations in existing methods that focus solely on subject identity, neglecting crucial visual storytelling elements like spatial composition, background evolution, and narrative pacing. This project benefits researchers and artists by enabling state-of-the-art diffusion models to produce visually diverse yet coherent storyboards without requiring architectural modifications or fine-tuning.

How It Works

Story2Board employs a lightweight, two-component consistency framework. Latent Panel Anchoring (LPA) stabilizes character identity across panels by reusing a shared reference latent. Reciprocal Attention Value Mixing (RAVM) enhances cross-panel coherence by softly blending visual features between token pairs exhibiting strong reciprocal attention. This approach leverages existing diffusion models, converting free-form stories into grounded panel-level prompts using an off-the-shelf LLM, thereby improving narrative engagement and visual consistency.

Quick Start & Requirements

• Platform Support: Officially supports Linux with Python 3.12 and CUDA 12.x. Windows/macOS are not officially tested but may work with requirements_all_platforms.txt. A Conda environment with Python 3.12 is recommended.
• Installation: Clone the repo, create/activate a Conda environment (conda create -n story2board python=3.12, conda activate story2board), and install dependencies (pip install -r requirements.txt). PyTorch should be installed first if a specific CUDA build is needed.
• Running: Use `main