Neural Doodle is an archived creative AI project that implements semantic style transfer with deep neural networks. It can transform simple annotated sketches into detailed images by borrowing visual characteristics from reference artwork or photographs. Semantic maps identify corresponding regions so style patches can be transferred to appropriate parts of a generated image. The software also performs conventional style transfer when semantic annotations are omitted. Additional modes support image analogy and standalone texture synthesis from source images. Users can adjust style strength, semantic weighting, smoothing, output resolution, and other parameters while running the tool on a CPU or CUDA-enabled GPU.
Features
- Semantic neural style transfer
- Annotated doodle-to-image generation
- Image analogy synthesis
- Standalone texture generation
- Adjustable style and semantic weights
- CPU, CUDA GPU, and Docker support