Transforms photos into Renaissance masterpieces, Van Gogh canvases, or Impressionist works, on demand.
Neural style transfer has existed since 2015, but early GAN-based approaches struggle with identity preservation and style intensity control. This project combines two paradigms: classical GAN-based neural style transfer for precise feature-level stylization, and Stable Diffusion with a custom LoRA adapter for more fluid, diffusion-based style application. The Van Gogh LoRA was fine-tuned on 400 curated paintings from the artist's distinct periods, capturing brush stroke texture, color palette, and compositional patterns that generic artistic prompts miss.
The GAN approach uses a pretrained VGG network to extract content and style features. Gram matrices capture texture statistics; gradient descent iteratively blends them in the output image.
400 Van Gogh paintings were curated, captioned, and used to fine-tune a LoRA adapter on Stable Diffusion. The adapter captures the artist's specific color vocabulary and brushstroke character.
With the Van Gogh LoRA loaded, img2img diffusion transforms input photos with a configurable strength parameter controlling how much of the original content is preserved.
Users choose between GAN (precise, feature-level) or Diffusion (fluid, impression-level) pipelines, and select from Renaissance, Impressionism, Expressionism, or custom styles.
Custom Van Gogh LoRA produces significantly more authentic stylization than generic "Van Gogh style" prompts
GAN pipeline preserves structural content better for portrait and architectural subjects
Diffusion pipeline produces more creative, atmospheric results for landscape subjects
Two-pipeline architecture gives users meaningful stylistic control