VidSketch: Hand-drawn Sketch-Driven Video Generation with Diffusion Control

TL;DR

VidSketch addresses the gap in generating video animations directly from hand-drawn sketches by introducing an abstraction-aware, diffusion-based framework. It combines Level-Based Sketch Control to adapt guidance strength to sketch abstraction with a TempSpatial Attention module to enforce spatiotemporal coherence across frames. The method demonstrates superior performance against baselines in both quantitative metrics and a user study, with ablations confirming the importance of LBSC and TSA. The approach lowers barriers for non-experts to create high-quality animated content across diverse styles, using only sketches and simple prompts.

Abstract

With the advancement of generative artificial intelligence, previous studies have achieved the task of generating aesthetic images from hand-drawn sketches, fulfilling the public's needs for drawing. However, these methods are limited to static images and lack the ability to control video animation generation using hand-drawn sketches. To address this gap, we propose VidSketch, the first method capable of generating high-quality video animations directly from any number of hand-drawn sketches and simple text prompts, bridging the divide between ordinary users and professional artists. Specifically, our method introduces a Level-Based Sketch Control Strategy to automatically adjust the guidance strength of sketches during the generation process, accommodating users with varying drawing skills. Furthermore, a TempSpatial Attention mechanism is designed to enhance the spatiotemporal consistency of generated video animations, significantly improving the coherence across frames. You can find more detailed cases on our official website.

Figures (15)

• Figure 1: Hand-drawn Sketch-Driven Video Generation. Our VidSketch empowers users of all skill levels to effortlessly create stunning, high-quality video animations using concise text prompt and hand-drawn sketch sequences of any level of abstraction.
• Figure 2: Video animations generated by our VidSketch. Our method generates video animation using any number of hand-drawn sketches (examples from top to bottom are guided by 1, 2, 4, and 5 sketches, respectively) and straightforward text prompt. This enables the creation of high-quality, spatiotemporal-consistent video animations, breaking barriers in the art profession. The figure showcases basic cases, and more high-quality examples can be found in our \ref{['fulu']}, supplementary materials, and on our https://csfufu.github.io/vid_sketch/.
• Figure 3: Pipeline of our VidSketch. During training, we use high-quality, small-scale video datasets categorized by type to train the TempSpatial Attention (TS-Attention) and Temporal Attention blocks, improving spatiotemporal consistency in video animations. During inference, users simply input a prompt and sketch sequences to generate tailored high-quality animations. Specifically, the first frame is generated using existing techniques, while the entire sketch sequence is processed by the Inflated T2I-Adapter mou2024t2i to extract information, which is injected into VDM's upsampling layers to guide video generation.
• Figure 4: The effectiveness of our Level-Based Sketch Control Strategy. We perform a quantitative analysis of the continuity, connectivity, and texture detail of sketches to automatically evaluate the abstraction level of hand-drawn sketche sequences. As the horizontal axis moves to the left, the abstraction level of the evaluated sketches gradually increases, and their corresponding abstraction scores become progressively larger.
• Figure 5: The details of the proposed TempSpatial Attention mechanism. We extract the $K/V$ tokens from the first frame, the second frame, and the $(i-1)$-th frame, and compute the attention mechanism using the query $Q$ from the $i$-th frame, which enhances the spatiotemporal consistency of the video animation.