Low-Cost Deep Learning for Automated Keyframe Interpolation in Sprite Animation
Sprite-based animation remains prevalent in video game development, but manual in-betweening---generating intermediate frames between keyframes---is labor-intensive. Existing video frame interpolation methods, optimized for photorealistic content, introduce smoothing artifacts when applied to the sharp edges, limited color palettes, and stylized motion characteristic of hand-drawn sprites. We present a deep learning approach for sprite in-betweening that adapts a U-Net encoder-decoder architecture with domain-specific modifications, including RGBA-aware processing, multi-scale flow refinement, and alpha-gated blending to preserve pixel-perfect sprite aesthetics. Following data-centric AI principles, we demonstrate that a few carefully curated sprite animation triplets---selected for diversity, quality, and representative edge cases---enable effective training without requiring massive generic datasets. Experimental evaluation using the FLIP perceptual metric shows that our method achieves a mean error of 0.258 on 42 test triplets, outperforming state-of-the-art video interpolation RIFE (0.313, 17.5\% worse), traditional optical flow Lucas-Kanade (0.449, 42.6\% worse), and general frame synthesis CyclicGen (0.603, 57.2\% worse). This study shows that quality-over-quantity data curation and domain-specific architectural adaptations yield better interpolation quality than general-purpose methods trained on orders of magnitude more data, validating the data-centric AI paradigm for specialized visual domains. This is particularly important for small-to-medium-sized game development studios, where budget and time constraints are critical.