I trained an anime image model in 2 days from scratch on 1 local GPU

Independent AI developer well9472 has achieved a significant milestone in efficient model training with Nanosaur-250M, a specialized text-to-image model for anime-style artwork. Unlike typical approaches that fine-tune existing models like Stable Diffusion, this project involved training both the VAE (Variational Autoencoder) and diffusion model components completely from scratch. The entire process took just 50 hours total—8 hours for the VAE and 42 hours for the diffusion model—using a single RTX Pro 6000 GPU and a dataset of 2 million anime illustrations.

The technical approach combines several recent innovations: DINOv3 serves as the encoder for the custom VAE, while Google's open-source Gemma3-270M model provides text encoding capabilities through the DeCo (Decoupled Contrastive Learning) framework. The model supports multiple high-resolution outputs including 832x1216, 896x1152, and 1024x1024 formats, accepting both tag-based and natural language captions. Well9472 has released not only the model weights but also complete training scripts, enabling other researchers to replicate the process or train similar models on different datasets.

This demonstration proves that creating capable domain-specific generative models no longer requires massive computational clusters or proprietary foundation models. The open-source release includes detailed documentation, inference scripts, and a technical report, making it a valuable resource for the AI research community interested in efficient training methodologies.