Story2Board

Story2Board: A Training‑Free Approach for Expressive Storyboard Generation

Project page: https://daviddinkevich.github.io/Story2Board
Paper (arXiv): https://arxiv.org/abs/2508.09983 Code: this repo

---

Abstract

We present Story2Board, a training-free framework for expressive storyboard generation from natural language. Existing methods narrowly focus on subject identity, overlooking key aspects of visual storytelling such as spatial composition, background evolution, and narrative pacing. To address this, we introduce a lightweight consistency framework composed of two components: Latent Panel Anchoring, which preserves a shared character reference across panels, and Reciprocal Attention Value Mixing (RAVM), which softly blends visual features between token pairs with strong reciprocal attention. Together, these mechanisms enhance coherence without architectural changes or fine-tuning, enabling state-of-the-art diffusion models to generate visually diverse yet consistent storyboards. We convert free-form stories into grounded panel-level prompts with an off-the-shelf LLM, and evaluate on a new Rich Storyboard Benchmark that measures layout diversity, background-grounded storytelling, and consistency. Qualitative/quantitative results and a user study show that Story2Board produces more dynamic, coherent, and narratively engaging storyboards than existing baselines.

---

⚙️ Installation

Platform Support

• ✅ Officially supported: Linux, Python 3.12, CUDA 12.x
• 💻 Windows / macOS: not officially tested, but you can try with the alternative requirements file: requirements_all_platforms.txt

We recommend a fresh Conda environment with Python 3.12.

---

1) Clone the repository

# 1) Clone the repository
git clone https://github.com/DavidDinkevich/Story2Board.git
cd Story2Board

# 2) Create and activate env
conda create -n story2board python=3.12
conda activate story2board

# 3) Install dependencies
pip install -r requirements.txt

Tip: If you want a specific CUDA build of PyTorch, install PyTorch first following the official instructions, then run pip install -r requirements.txt.

---

Quickstart

The entry point is main.py. The required arguments are:

• --subject – the main subject (e.g., “smiling boy”).
• --ref_panel_prompt – description of the reference (top) panel.
• --panel_prompts – one or more prompts for the remaining panel(s).
• --output_dir – where to save results.

Minimal skeleton:

python main.py   --subject "SUBJECT_NAME"   --ref_panel_prompt "REFERENCE_PANEL_TEXT"   --panel_prompts "PANEL_1_TEXT" "PANEL_2_TEXT" ...   --output_dir path/to/out

Concrete example

python main.py   --subject "fox with shimmering fur and glowing eyes"   --ref_panel_prompt "stepping onto a mossy stone path under twilight trees"   --panel_prompts     "bounding across a fallen tree over a mist-covered ravine glowing faintly with constellations"     "perched atop a broken archway of ancient stone, vines and silver moss hanging down, the twilight sky glowing behind him"     "watching a meteor shower from the edge of a luminous lake that reflects the stars perfectly"   --output_dir outputs

This will generate a storyboard where the top panel is the reference, and each bottom panel reuses the same character identity while varying the scene/action.

---

Outputs

• Generated images are written to --output_dir.
• The constructed, per-panel prompts are logged for reproducibility.

---

Method Overview (Very Brief)

• Latent Panel Anchoring: reuses a shared reference latent to stabilize identity across panels.
• RAVM (Reciprocal Attention Value Mixing): gently blends attention values between token pairs with strong reciprocal attention, preserving the model’s prior while improving cross-panel coherence.

---

Changelog

• 21 Aug 2025: Fixed missing line for Latent Panel Anchoring (LPA) in the released code (commit 34537e0).
  • Restores expected behavior described in the paper.

BibTeX

@article{dinkevich2025story2board,
  title={Story2Board: A Training-Free Approach for Expressive Storyboard Generation},
  author={Dinkevich, David and Levy, Matan and Avrahami, Omri and Samuel, Dvir and Lischinski, Dani},
  journal={arXiv preprint arXiv:2508.09983},
  year={2025}
}

Acknowledgements

This repository builds on the excellent open-source ecosystems of PyTorch and Hugging Face Diffusers, and uses FLUX.1-dev weights as the base T2I model.

---

License

See LICENSE in this repository.