Image & Video

Redefining Art in 2026: From Sketch-Based Models to Full Image Generation

A new UNET-based AI creates high-res art using only personal photos and public domain images, avoiding copyright issues.

Deep Dive

Developer Jason Juan has unveiled a novel approach to AI image generation with his custom-built model, codenamed 'Milestone / Jason 10M Model'. The system is based on a UNET neural network architecture, which learns to transform noise into coherent images by recognizing patterns like shapes and textures. The key innovation lies in its training data: instead of scraping the internet, the model was trained exclusively on Juan's personal photographs and public domain images. This creates a fully traceable lineage for all outputs, directly addressing copyright and licensing concerns that plague large-scale models like Stable Diffusion or DALL-E.

The project also features a two-stage training process. First, a smaller 256x256 'sketch model' was trained to recognize foundational objects like chairs and tables. This model then informs a larger system capable of generating images at a native 1024x1024 pixel resolution. The results challenge the prevailing assumption that high-quality AI art requires massive, opaque datasets and cloud-scale compute. Juan's work proves that with careful architectural design and optimization, impressive results can be achieved with a controlled, legally-sourced dataset.

Positioned as an alternative path for responsible AI development, the 'Milestone' model emphasizes principles of data ownership, reproducibility, and independence from proprietary datasets. Its development aligns with emerging regulatory considerations, such as California's Assembly Bill 2013, by prioritizing transparency in dataset composition. While not yet publicly released, the project signals a shift towards more ethical and controllable creative AI tools, offering a blueprint for developers who prioritize copyright compliance and artistic intent over sheer scale.

Key Points
  • Trained exclusively on personal photos and public domain images, ensuring 100% copyright-compliant and traceable outputs.
  • Uses a two-stage UNET architecture with a 256x256 sketch model to learn basic shapes before generating 1024x1024 images.
  • Demonstrates high-quality image generation is possible without massive internet-scale datasets, emphasizing data ownership and transparency.

Why It Matters

Offers a blueprint for ethical, copyright-compliant AI art, reducing legal risks for commercial use and empowering individual creators.