In November of 2024, the first models made for the NovelAI's Image Generator have been released for public.<a target="_blank" rel="ugc" href="https://blog.novelai.net/novelai-diffusion-v1-weights-release-en-e40d11e16bd5">https://blog.novelai.net/novelai-diffusion-v1-weights-release-en-e40d11e16bd5</a>And someone in HuggingFace have converted the released versions from Unet into Safetensors, and those are the versions I upload to this project.<a target="_blank" rel="ugc" href="https://huggingface.co/TechnoByte/nai-furry-anime-safetensors">https://huggingface.co/TechnoByte/nai-furry-anime-safetensors</a>

NovelAI, NAI Diffusion V1

novelaiNAIDiffusion_v10Curated.safetensors

v1.0 Curated

v1.3 Furry

v1.0 Full

<a target="_blank" rel="noopener noreferrer nofollow" href="https://blog.novelai.net/novelai-diffusion-v1-weights-release-en-e40d11e16bd5">NovelAI Diffusion V1 Weights Release (EN)</a>As the field of image generation progresses, we have decided to release the weights of our first generation image generation models to the public, for the purposes of research, personal use and historical preservation. Of course, they will also remain available for use on our website, but now users who want to play with them locally can freely download and use these models.The specific models we are releasing are:• <a target="_blank" rel="noopener ugc nofollow" href="https://huggingface.co/NovelAI/nai-anime-v1-curated">NovelAI Diffusion Anime V1 (Curated)</a> • <a target="_blank" rel="noopener ugc nofollow" href="https://huggingface.co/NovelAI/nai-anime-v1-full">NovelAI Diffusion Anime V1 (Full)</a> • <a target="_blank" rel="noopener ugc nofollow" href="https://huggingface.co/NovelAI/nai-furry-beta-v1.3">NovelAI Diffusion Furry (Beta V1.3)</a>The “curated” anime model was trained on a smaller, higher quality set of images, has better knowledge of pop culture and higher image quality. The “full” anime model was trained on a more comprehensive set of images, making it more flexible. Finally, we are also releasing the weights of our old furry model.While we are taking a trip down the memory lane, here are some of our relevant blog posts from back in the day, when we had just released these models on our website:• <a target="_blank" rel="noopener" href="https://novelai.medium.com/image-generation-announcement-807b3cf0afec">NovelAI Diffusion Image Generation Announcement Blog Post</a> • <a target="_blank" rel="noopener ugc nofollow" href="https://blog.novelai.net/novelai-improvements-on-stable-diffusion-e10d38db82ac">NovelAI Improvements on Stable Diffusion</a> • <a target="_blank" rel="noopener ugc nofollow" href="https://blog.novelai.net/the-magic-behind-novelaidiffusion-b4797e0d27b2">The Magic behind NovelAIDiffusion</a> • <a target="_blank" rel="noopener ugc nofollow" href="https://blog.novelai.net/novelai-aspect-ratio-bucketing-source-code-release-mit-licensed-c33fdd25e4ad">NovelAI Aspect Ratio Bucketing Source Code Release (MIT Licensed)</a>As written in the second of those blog articles, our models were trained using the output of the penultimate layer of the CLIP text encoder, so when using these models, make sure to set your image generation UI to use the second to last CLIP layer rather than the last one.The models are released under a dual license, with both terms of the original CreativeML Open RAIL-M license and the Creative Commons BY-NC-SA 4.0 license applying to them. This allows non-commercial: use, redistribution and the creation of derivative works, as long as they are attributed properly and distributed under the same terms.Of course, these models are all a bit dated, so if you just want to have fun with a state of the art anime or furry image generation model, make sure to check out NovelAI V3 on <a target="_blank" rel="noopener ugc nofollow" href="https://novelai.net/">our website</a> and stay tuned for future updates!To reiterate, the weight release is only applicable to the earliest iterations of our NovelAI Diffusion models dating back to October (Anime Diffusion Curated / Full) and November (Furry Diffusion V1.3) 2022

This is our oldest anime model, trained on a curated dataset. We are now making the weights publicly available for research and personal use. To keep the release compact, it only contains the Unet and VAE components. If you require the other components, you can copy them into the folder from the original Stable Diffusion 1.4 release.

The model was trained using the text embeddings produced by CLIP's penultimate layer, so, depending on your inference software, you will need to set "CLIP skip" to 2.

Files with SHA256: 2c29e4d408e7d9030d03abad1370fcaa0c8f43c992b3a6b6afbdda35d12454e4