UCSC-VLAA/story-iter

[ICLR 2026] A Training-free Iterative Framework for Long Story Visualization

What it solves

Story-Iter addresses the difficulty of maintaining semantic consistency and high-quality interactions when generating long sequences of images for a story (up to 100 frames). Traditional text-to-image models often struggle to keep characters and settings consistent across many frames without expensive training or relying on a few fixed reference images.

How it works

It uses a training-free iterative paradigm that goes beyond the standard denoising steps of diffusion models. The process works in rounds:

  1. Initialization: It first generates an initial set of images based solely on the story's text prompts.
  2. Iteration: In each subsequent round, it uses all images from the previous round as references.
  3. Global Reference Cross-Attention (GRCA): A plug-and-play module is inserted into Stable Diffusion to aggregate information from all reference frames using global embeddings, ensuring that the visual context is maintained across the entire sequence.

Who it’s for

This tool is designed for creators and researchers interested in automated story visualization, comic generation, or creating long-form visual narratives where character and scene consistency is critical.

Highlights

  • Training-free: Does not require additional model training to achieve consistency.
  • Long-sequence support: Capable of visualizing stories up to 100 frames.
  • Flexible styles: Supports various output styles including comic, film, and realistic.
  • ControlNet integration: Supports openPose skeletons as control signals for precise character posing.
  • Fast execution: Includes a fast version (via LCM) that can process a 100-frame story over 10 iterations in approximately 20 minutes.