We introduce Calligrapher, a novel diffusion-based framework that innovatively integrates advanced text customization with artistic typography for digital calligraphy and design applications. Addressing the challenges of precise style control and data dependency in typographic customization, our framework incorporates three key technical contributions. First, we develop a self-distillation mechanism that leverages the pre-trained text-to-image generative model itself alongside the large language model to automatically construct a style-centric typography benchmark. Second, we introduce a localized style injection framework via a trainable style encoder, which comprises both Qformer and linear layers, to extract robust style features from reference images. An in-context generation mechanism is also employed to directly embed reference images into the denoising process, further enhancing the refined alignment of target styles. Extensive quantitative and qualitative evaluations across diverse fonts and design contexts confirm Calligrapher's accurate reproduction of intricate stylistic details and precise glyph positioning. By automating high-quality, visually consistent typography, Calligrapher surpasses traditional models, empowering creative practitioners in digital art, branding, and contextual typographic design.
Training framework of Calligrapher, demonstrating the integration of localized style injection and diffusion-based learning. The framework processes masked images through a Variational Auto-Encoder (VAE) to obtain latent representations, concatenated with mask and noise latents. A style encoder comprising a visual encoder, Qformer, and linear layers is designed to extract style-related features from the reference style image, while text embeddings (e.g., "gic" in the case) modulate the denoising transformer. In the denoising block, style attention predicted from the style features replaces the original cross-attention, injecting style embeddings with the denoiser's query to enable granular typographic control in the latent space. The model is optimized under the flow-matching learning objective with the self-distillation typography dataset.
Qualitative results of Calligrapher under various settings. We demonstrate text customization results respectively under settings of (a) self-reference, (b) cross-reference, and (c) non-text reference. Reference-based image generation results are also incorporated in (d).
