Title: Pinterest Canvas: Large-Scale Image Generation at Pinterest

Abstract: Although contemporary image generation models exhibit impressive versatility across numerous tasks, this broad flexibility often complicates control through prompting or basic inference adjustments. Consequently, these general-purpose models are frequently ill-suited for applications demanding strict product standards. To address this challenge, we present Pinterest Canvas, a robust large-scale image generation infrastructure designed specifically to facilitate image editing and enhancement within the Pinterest ecosystem.

Our approach begins by training a foundational diffusion model on a heterogeneous, multimodal dataset, endowing it with extensive image-editing capabilities. Rather than attempting to force a single generic model to manage every downstream objective, we employ a strategy of rapidly fine-tuning specialized variants of this base architecture on task-specific datasets. This method yields dedicated models tailored to individual use cases. This paper outlines the core components of the Canvas system and shares our established best practices regarding dataset curation, training protocols, and inference procedures.

Through case studies focusing on background enhancement and aspect-ratio outpainting, we illustrate how the system addresses distinct product requirements. Online A/B testing results indicate that these enhancements drive substantial user engagement, achieving lifts of 18.0% and 12.5%, respectively. Furthermore, evaluations involving human raters confirm that our models surpass third-party alternatives in these specific tasks. Finally, we demonstrate the system’s broader applicability by showcasing other Canvas variants, such as multi-image scene synthesis and image-to-video generation, proving that our methodology effectively generalizes across a wide spectrum of potential downstream applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC