New uncensored image model! Nanosaur 2 full release!
retained text
Hello everyone. I have to preface this by saying this is NOT my model, though I'm very familiar with the person who made it. I can with full confidence tell you that for any model smaller than anima, this is the best one out there. This is a 670m parameter image model trained on all of e6/danbooru, 14 million images. Takes natural language OR tags, your choice. It comes packaged with a new vae, a 270m parameter text encoder, and a 670m dit. **This model can do a very large amount of concepts, artist styles, and characters.** That being said, it's still a 670m parameter image model, so don't expect it to move mountains. Compared to SDXL model though, it wins a lot of the time. By default, the actual model is 1.3gb, and with the text encoder and vae loaded, it rises to 2gb. Too big for you? there are int8 and w4a8 quants that bring it down to as low as 388mb for the dit. So this can be run on a toaster if you're looking for that Links: Base Repo: [link](https://huggingface.co/well9472/Nanosaur2-670M) int8: [link](https://huggingface.co/bertbobson/Nanosaur2-670M-INT8-ConvRot/blob/main/nanosaur2_int8_conv.safetensors) w4a8 (not recommended): [link](https://huggingface.co/bertbobson/Nanosaur2-670M-INT8-ConvRot/blob/main/nanosaur2_w4a8.safetensors) Also, this project is looking for funding for a bigger version. If you'd like to help out with that or discuss terms, please reach out to **metal63** on discord.
document fetched 2026-09-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wqh0ti/new_uncensored_image_model_nanosaur_2_full_release