IdentityStory: Taming Your Identity-Preserving Generator for Human-Centric Story Generation

TL;DR

IdentityStory tackles human-centric story generation by decoupling identity extraction and identity injection. It introduces Iterative Identity Discovery, which uses SVD-based filtering to obtain cohesive identity embeddings from multiple character images, and Re-denoising Identity Injection, which combines a general text-aligned generator for layout with an identity-preserving generator for identity injection via re-denoising guided by character masks. Evaluations on ConsiStory-Human show superior face consistency and competitive text alignment, with strong evidence from quantitative metrics and human judgments. The approach enables multi-character, infinite-length storytelling and dynamic composition, offering a flexible, practical path toward human-centric visual narratives.

Abstract

Recent visual generative models enable story generation with consistent characters from text, but human-centric story generation faces additional challenges, such as maintaining detailed and diverse human face consistency and coordinating multiple characters across different images. This paper presents IdentityStory, a framework for human-centric story generation that ensures consistent character identity across multiple sequential images. By taming identity-preserving generators, the framework features two key components: Iterative Identity Discovery, which extracts cohesive character identities, and Re-denoising Identity Injection, which re-denoises images to inject identities while preserving desired context. Experiments on the ConsiStory-Human benchmark demonstrate that IdentityStory outperforms existing methods, particularly in face consistency, and supports multi-character combinations. The framework also shows strong potential for applications such as infinite-length story generation and dynamic character composition.

Figures (15)

• Figure 1: Human-centric story generation. Our IdentityStory can solely rely on text to generate a series of images that consistently depict human characters and faithfully align with text prompts, outperforming the state-of-the-art method StoryDiffusion. Zoom in for better view.
• Figure 2: Identity embeddings of identity-preserving generators. Here we use PhotoMaker li2024photomaker as an example: (a) We extract embeddings of three identities using Photomaker's image encoder and visualize them in 2D with t-SNE van2008visualizing, observing that embeddings of different identities are highly distinguishable. (b) We collect 20 character descriptions and extract identity embeddings using the naive averaging approach and our method. Then, each embedding generates 15 images for computing pairwise face similarity with ArcFace deng2019arcface, showing our method performs better.
• Figure 3: Text alignment degradation of identity-preserving generators. We collect 1,000 human-centric text prompts to compare the text alignment performance of vanilla SDXL podell2023sdxl, PhotoMaker li2024photomaker, and InstantID wang2024instantid. We present (a) qualitative and (b) quantitative results with the average CLIP score (CLIP-T) radford2021learning. For the identity-preserving generators, we use SDXL as the base model and generate a human image as the reference.
• Figure 4: Pipeline of IdentityStory. This framework consists of two key techniques, including (a) Iterative Identity Discovery (Section \ref{['sec:techa']}), which utilizes Singular Value Decomposition (SVD) to iteratively filter out low-relevance embeddings to extract cohesive identities, and (b) Re-denoising Identity Injection (Section \ref{['sec:techb']}), which uses identity-preserving generators to inject extracted identities based on the noisy cached images and character layouts produced from general generators.
• Figure 5: Desgin choices of Re-denoising Identity Injection. (a) We start from a sweet-spot timestep ($t^\prime=40$) to re-denoise, balancing image harmony and identity fidelity (blue frame). (b) We develop a progressive masking strategy to effectively diminish artifacts (red frame).