well9472/Nanosaur2-670M

This artifact is a 670M parameter text-to-image diffusion transformer designed for research. The author explicitly states the goal is to test minimal compute limits rather than compete with fully trained commercial systems. It generates illustrations using a frozen Gemma text encoder and a custom DINOv2-based VAE. The model was trained for a total cost of approximately $600. Integration requires installing a custom node into the ComfyUI environment and placing three specific model files in designated folders. A provided workflow file allows users to load the configuration directly into the interface. Inference relies on Euler simple sampling with 50 steps and a CLIP guidance scale of 4. Users can employ tag-based prompts or natural language, often using quality tags to adjust output preferences. Potential users should verify that their hardware matches the specific inference requirements documented by the publisher. The author notes that the current version does not match the character knowledge of larger commercial models. There is uncertainty regarding long-term stability since the author states this model was not intended for production use. Further development, including a larger 2B version, depends on securing additional compute funding.