Abstract
This publication discloses Style-Preserved Inference Flow (SPIF), a system for real-time style-locked image generation that maintains consistent artistic style across frames at interactive rates. The core innovation is structural model pruning guided by style-consistency metrics rather than general image quality, producing compact single-purpose models that generate images in a specific artistic style at 60+ frames per second. Additional disclosed techniques include latent-space frame coherence (bootstrapping sequential frames from previous latent states, analogous to dirty-rectangle optimization in game rendering), composable latent object descriptors for real-time scene manipulation, and voice-driven continuous generation loops. Working prototypes demonstrate 3.35 FPS full-pipeline generation at 1024×1024 and 12 FPS latent-only processing on consumer hardware, with clear optimization paths to 60 FPS targets. Prior art in the inventor's game development work (1984–present) establishes the cross-domain pattern matching underlying these innovations.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Sharp, Douglass W., "Style-Preserved Inference Flow (SPIF): Real-Time Style-Locked Image Generation Through Model Pruning and Latent Space Optimization", Technical Disclosure Commons, (February 11, 2026)
https://www.tdcommons.org/dpubs_series/9323