Nanosaur 2
A lightweight open-source image generation model based on the DiT (Diffusion Transformer) architecture, with 670M parameters, a 270M parameter text encoder, and a new VAE. It supports both natural language and tag-based inputs, and was trained on an anime-style dataset.
Timeline (last 90 days)
Nanosaur 2 is an open-source image generation model with 670 million (670M) parameters.
Nanosaur 2 was officially released in the Reddit community.
The core of Nanosaur 2 is a 670M-parameter DiT (Diffusion Transformer) architecture, paired with a brand-new VAE and a 270M-parameter text encoder.
Nanosaur 2 was trained on the full e6 and danbooru datasets, totaling approximately 14 million images.
Nanosaur 2 supports both natural language descriptions and tags as input methods.
According to the publisher's description, Nanosaur 2 outperforms the SDXL model in many comparisons, despite having far fewer parameters than SDXL.
The Nanosaur 2 model itself is approximately 1.3 GB, and the total size is approximately 2 GB after loading the text encoder and VAE.
Nanosaur 2 offers an int8 quantized version and a w4a8 quantized version, with the w4a8 version compressing the DiT component to as low as approximately 388 MB.
The publisher has labeled the w4a8 quantized version as 'not recommended' due to potentially significant image quality degradation.
Nanosaur 2 is labeled as uncensored and has no built-in content filtering restrictions.
2 more timeline events
All Facts (11)
The publisher stated that Nanosaur 2 is not their own work, but rated it as the best model currently available among any model smaller than anima.
50%UnverifiedNanosaur 2 is an open-source image generation model with 670 million (670M) parameters.
50%UnverifiedThe core of Nanosaur 2 is a 670M-parameter DiT (Diffusion Transformer) architecture, paired with a brand-new VAE and a 270M-parameter text encoder.
50%UnverifiedNanosaur 2 was trained on the full e6 and danbooru datasets, totaling approximately 14 million images.
50%UnverifiedNanosaur 2 supports both natural language descriptions and tags as input methods.
50%UnverifiedAccording to the publisher's description, Nanosaur 2 outperforms the SDXL model in many comparisons, despite having far fewer parameters than SDXL.
50%UnverifiedThe Nanosaur 2 model itself is approximately 1.3 GB, and the total size is approximately 2 GB after loading the text encoder and VAE.
50%UnverifiedNanosaur 2 offers an int8 quantized version and a w4a8 quantized version, with the w4a8 version compressing the DiT component to as low as approximately 388 MB.
50%UnverifiedThe publisher has labeled the w4a8 quantized version as 'not recommended' due to potentially significant image quality degradation.
50%UnverifiedNanosaur 2 is labeled as uncensored and has no built-in content filtering restrictions.
50%UnverifiedNanosaur 2 was officially released in the Reddit community.
50%