VideoSketcher: Sequential Sketch Generation Using Video Model Priors

A research team led by Hui Ren has introduced VideoSketcher, a method that adapts pretrained text-to-video diffusion models to generate the sequential process of sketching, treating drawing not as a static image but as a temporally coherent sequence of strokes [1][2]. The work addresses a longstanding gap in generative modeling. Most systems produce sketches as finished images, ignoring the progressive, stroke-by-stroke nature of drawing that is central to creative exploration [2]. Prior attempts to model this sequential structure either depended on large, human-drawn datasets with limited diversity or used large language models to issue drawing instructions, a strategy that often sacrificed visual fidelity [2]. VideoSketcher instead combines the complementary strengths of large language models and video diffusion models: the language model acts as a semantic planner that decomposes a concept into step-by-step instructions, while the video diffusion model serves as a renderer that translates those instructions into temporally coherent sketch sequences [2]. The framework relies on a two-stage fine-tuning strategy that decouples temporal structure from visual appearance. Stroke ordering is learned from synthetic shape compositions, and style is distilled from as few as seven hand-drawn examples [2]. Despite this minimal supervision, the method can generate diverse, high-quality sequential sketches that follow specified drawing orders [2]. The approach also extends to brush style control and autoregressive generation, opening possibilities for artistic applications [2]. The paper was submitted to the arXiv preprint repository on February 17, 2026, and revised on June 18, 2026 [1]. arXiv, operated by Cornell University, provides rapid dissemination of research findings at no cost to readers and submitters [11]. The repository hosts a framework called arXivLabs, which enables community collaborators to develop and share experimental tools directly on article pages, such as bibliographic explorers and code finders [10][11]. These tools are built under guidelines that require partners to uphold values of openness, community, excellence, and user data privacy [10].