Novel AI Diffusion - Anime / Furry (beta) oriented Stable Diffusion made easy

Currently reading:
Novel AI Diffusion - Anime / Furry (beta) oriented Stable Diffusion made easy

Xeraser

Professional shade thrower.
comfy friend
Joined
Apr 4, 2022
Messages
102
Reaction score
68
comfy coins
💠353,949
Novel AI Diffusion is a Stable Diffusion service/setup that's anime-oriented (with a furry model in beta if you're into that) and is braindead-easy to use.
The service requires a subscription of 15$/month minimum, with a recommended 25$/month plan to generate an unlimited amount of images at 640x640, 28 steps maximum. While it may seem somewhat steep you also need to remember that you get access to Novel AI's other features (arguably the main ones) such as the Novel part of the service (which is surprisingly fun to use even if you have zero interest in writing a story) and the excellent TTS. With the 25$/month tier you get 10,000 generation tokens that you can use to pay for larger (up to 1024x1024, 50 steps) images. Said large images cost about 25 tokens at max settings, 20 if you use 40 steps (which seems to be the euler_ancestral sampler's soft cap). Extra currency costs 3.79/6.49/10.99 USD for 2000/5000/10,000 tokens respectively.
Some of their hardware has been quarantined after a very serious breach so the service isn't running as fast nor as well as it should right now - but on a normal day it takes 5 seconds flat to generate an image at the free settings (28 steps, up to 640x640, usually 512x768 portrait or 768x512 landscape resolutions) which makes constant prompt-tweaking much less frustrating than self-hosted SD.

"But Xeraser, I don't want to pay/I'm too cheap/I just bought a 3090 Ti from a failed crypto farm!"

You're in luck - mostly.
Novel AI's models, frontend and backend (yes, you can host the SITE ITSELF on your hardware and get a nearly 1:1 experience) have been leaked, all of which you can find in the guides below.

Requires making a SD setup: NAI leaked model (full leak, 50+ GB but not everything is required) to use with Stable Diffusion's webui: https://rentry.org/sdg_FAQ#how-do-i-setup-the-leaked-novelai-model
Requires NO SD setup, runs on Windows/Linux, REQUIRES an Nvidia card with 8GB+ VRAM: NAIFU ("full" model + frontend + backend): https://rentry.org/sdg_FAQ#naifu-novelai-model-backend-frontend (you can use the other models from the full leak above with this or other SD models entirely, read the guide at the bottom of the section)

xformers - do you need them? Possibly, possibly not. They WILL affect the final result, possibly for the worse. If in doubt just don't bother.

IMPORTANT - HERE'S WHY YOU MIGHT WANT TO PAY FOR THE SERVICE:
  • Speed. On a 1080 Ti the default settings that NAI uses take nearly 4 times longer to generate an image or roughly 18.x seconds VS NAI's 5 seconds flat. On maximum setting (1024x1024, 50 steps) it takes over two full minutes vs NAI's 15 seconds average). With how RNG-y Stable Diffusion is the extra time required will make it significantly less fun and more frustrating to mess around with.
  • Power/Time/Price ratio. Considering the example above it might end up being cheaper vs your electricity bill depending on where you live (especially Europe)
  • A significant one: constant updates and fixes. With NAI proper you'll get constant under-the-hood updates, possibly new/updated models (the furry one will be updated to get it out of beta), QoL things (for example you can now drag and drop a previously generated image to copy every setting or every setting PLUS the seed automatically), support from the devs and more. Possibly none of these will come to NAIFU (leaked frontend+backend+model) unless another huge leak/breach happens (remember, a 0-day was used to get the first leak, don't think that will happen again anytime soon)
    While some of these MAY be possible (mostly the QoL updates) you're still at the mercy of 4chan anons.
  • Consistency and overall quality: NAIFU results seem to be 10-20% worse than NAI proper (might be even worse with the leak + webui) and this most likely has to do with the above. Identical settings and seeds will not produce 1:1 results between them: NAI outputs might look worse in NAIFU, NAIFU outputs might look better in NAI, etc. This is probably due to NAI constantly fucking with/tweaking their sampler as even NAI to NAI results seem to vary slightly on a daily basis when using identical settings and seeds (but still closer than NAI vs NAIFU)

How do I use the furry model with the leak? Is it even included?
I don't know and I don't care to know until it can give me actual Imp Midna.
 
Last edited:
NAI was having server issues for the past 4-5 days and they were particularly bad today. They moved part of their backend to a different provider and the image stuff will be handled by a new backend service. It now takes 3 seconds flat instead of 5 to generate an image.

3EA6D3E.png
 
Well it works well enough for me so I am very happy considering I don't have to burn my gpu making them locally and being entirely free(as in cost)
Burning GPUs is a bit of a meme, just preemptively set the fan to 100% in afterburner or something and you'll be fine
Maybe undervolt if you can or set power limits, etc etc
 
Update to the update: there was a stealth update, hands are coming out much better than before. Not perfect but better.
 
ATTENTION EVERYONE
NAI has been superseded for months so uh don't sub anymore lol just use paperspace pro if you don't have an 8gb+ gpu
 
/pub/ ~ public channel
Help Users
  • N (Guest) Notxeraser:
    Oh yeah it works there now, I didn't use awave because it doesn't really have any useful formats for what I'm trying to do right now (though at this point I might as well get to work and Frankenstein an XS/XF version of this library)
    Quote Link
  • N (Guest) Notxeraser:
    Wasn't in my plans but might as well
    Quote Link
  • N (Guest) Notxeraser:
    I did try to convert the version from archive.org to something more usable like GIGA (played by g-player or by exs24) but that went horribly wrong
    Quote Link
  • N (Guest) Notxeraser:
    G-player under logic (at least 2.0) is buggy as all hell
    Quote Link
  • A (Guest) anon:
    If it works for you, I guess it's fine. It really did not look to me like there was anything wrong with the ISO, after the track mode change.
    Quote Link
  • N (Guest) Notxeraser:
    Oh no the iso is fine, thanks for that
    Quote Link
  • A (Guest) anon:
    If the archive.org version is the one I think it is, it's exactly the same as the rutracker one.
    Quote Link
  • N (Guest) Notxeraser:
    Guess it was just a very sloppy conversion job to get it to fit on an AKAI CD
    Quote Link
  • A (Guest) anon:
    And it looks like it's also the same as the magesy one, too.
    Quote Link
  • N (Guest) Notxeraser:
    anon said:
    If the archive.org version is the one I think it is, it's exactly the same as the rutracker one.
    it's half the size
    Quote Link
  • N (Guest) Notxeraser:
    Seems to have less programs too
    Quote Link
  • A (Guest) anon:
    I don't know, I have three or four versions I got from different places.
    Quote Link
  • N (Guest) Notxeraser:
    The archive.org one should be a little over 200mb
    Quote Link
  • A (Guest) anon:
    You are right that the archive.org one was smaller. It also was an ISO, and not a bin+cue.
    Quote Link
  • A (Guest) anon:
    Yup.
    Quote Link
  • N (Guest) Notxeraser:
    Yeah forgot to mention it was an iso
    Quote Link
  • A (Guest) anon:
    Anyway, the AKAI stuff itself does not look particularly wrong. If anything, I guess something went wrong when ripping it; but it's really strange that it would result in a cue file with the wrong track type.
    Quote Link
  • N (Guest) Notxeraser:
    What's Mode 2 used for anyway? Never gave it a single thought
    Quote Link
  • N (Guest) Notxeraser:
    with that said I haven't burned a CD in well over a decade
    Quote Link
  • N (Guest) Notxeraser:
    I think the last time I did it was to burn a copy of SSBB to a DL DVD for some reason
    Quote Link
  • A (Guest) anon:
    I'm not sure. I think it's something to do with CD-XA.
    Quote Link
  • A (Guest) anon:
    Essentially, mixing CD-ROM and CD-DA tracks in a single CD.
    Quote Link
  • A (Guest) anon:
    Not sure what mode 2 would have to do with skipping the first 8 bytes, however.
    Quote Link
  • OPdbx @ OPdbx:
    how to embed video file in comment? not just name of file...
    Quote Link
  • OPdbx @ OPdbx:
    not sure if my chats are being sent. is red normal?
    Quote Link
    OPdbx @ OPdbx: not sure if my chats are being sent. is red normal?
    Back
    Top