Western Animation Diffusion
Comicbook and Western Animation Style Model
Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕.
A ❤️, a kind comment or a review is greatly appreciated.
Purpose of this model
Train character loras where the dataset is mostly made of cartoon screencaps or comicbooks, allowing less style transfer and less overfitting.
Add variety to mixes.
Have an alternative to anime models when it comes to western stuff.
Not for western style lora training. Pretty good when used with western style loras (thanks @AnteMaxx), but needs weight adjusting, as it's already closer to the lora style probably.
Suggested settings
Set the ETA Noise Seed Delta (ENSD) to 31337
Set CLIP Skip to 2
DISABLE face restore. It's terrible, never use it
Use negative prompts and embeddings that don't ruin the style
Use AnimeVideo or Foolhardy as upscalers in highres fix
Use ADetailer for far away shots or full body images to avoid blurred faces
Brief history
This was requested by a supporter and I also wanted to see if I was capable of doing it. It was a funny little project.
Description
FAQ
Comments (33)
What does this do?
Set the ETA Noise Seed Delta (ENSD) to 31337
It adds 31,337 to your seed number every time you generate. It doesn't really do anything unless the picture generation info you're trying to copy also had their ENSD set to 31337.
Nothing, really. If you set your seed to 100, and your ENSD to 50, your resulting seed would be 150. The problem is that if you forget to clear out your ENSD. If your ENSD was set to 31337 and you found an image on CivitAI that you wanted to duplicate, you could copy and paste the exact same prompt and settings, including the seed, but then never generate that image, because you told SD to always add "31337" to the seed behind your back.
ENSD is one of those voodoo things that people used to think made some sort of difference in the image generation but all it does is change every seed by an exact amount. Some people use it as a "personal PIN" which means they can share their image on this site, with their prompts and settings, but unless we know their ENSD, we will never be able to generate that exact image.
@Devias original it was to reproduce NovelAI arts using their leaked model. Most people around in that period have that setting enabled
Even though I'm new, I pored over (almost) every nook & cranny of Automatic1111, so I know "what it is" and what it does, but I didn't know the "why" or "how" it came to be. I thought it was just some weird tradition that all SD users abide by to honor the AI founding fathers a thousand years ago or some such. So I'm like, okay whatever, when in Rome I guess... (Like "Hello, world." in coding.) So thanks for the background story & history, veterans.
@CaL9Mi7Y yeah basically since every model creator was using it, people kept using it to reproduce their stuff, and so why even change it. Once you have the setting there it just stays. Plus now Auto1111 webui has the setting imported from pnginfo. Back in the day you had to set it manually, so it stayed.
I've read back when NAI was released on ENSD, didn't understand it back then. Reading below now I still have no idea what it is. It's specific settings but abbreviated to the point it makes little sense. You can go search high and low in the settings and not know what the heck is ESND. Would be great if they just said what the settings were but guess ESND is a faster way of typing those specific settings out. Unfortunately it leaves a lot of us in the dark of what needs to be set where.
@spoffninja it's just a thing that makes the seed behave differently. It doesn't improve quality in any way and you don't need to use it unless you want to reproduce exactly my examples. It's not an important setting.
Can you please link to the upscale you suggested we use?
You can find them in here: https://upscale.wiki/wiki/Model_Database
@mramer723 Think I might be miss understanding here. But Foolhardy seems to be an uploader and not a specific upscaler? Does this mean ALL models that Foolhardy uploaded are required to replicate the upscaling. Is it a certain Foolhardy model that is used, if so which one?
As for AnimeVideo, can't see any model on that page that is called that, Is this a slang name that is refering to a specific model on that page? Was it deleted or maybe renamed?
@spoffninja I had AnimeVideo in my auto1111 installation by default. Anyway you don't really need foolhardy or animevideo. Most gans behave similarly with 0.4 denoising. Unless you use Latent, you'll get similar results.
Thank you for training this model! <3
It looks great and I can't wait to play with it in order to enhance the "comic" style of my model.
Of course I'll put the link to your model and I'll suggest people to buy you a coffee or support you on your patreon! Cheers!
Please go for it. I made this specifically so that people who do good mixes have an extra tool that's an alternative to anime models.
Just wanted to leave a word of encouragement. I like your models the best, Lykon.
Noise is all I get when I use this model - all other models work as normal, but whenever I use this one, despite following your settings to the letter, it comes out as an image of just noise. I'm sure this is a user error on my part, given this doesn't appear to have happened to anyone else. Any tips?
I see other reviewers using it just fine
:(
@Lykon I'm fairly sure that the file was corrupted somehow when I was downloading it, rather than there being anything wrong with your model itself. Re-downloading now - I very much look forward to using it!
Looks great! Can we get a pruned version?
i tried to make but face was deformed sometimes eyes too
my bet is that you're generating low res pics or far away ones and not using adetailer or highres fix.
This is the best model ever. Nothing comes close to this when it comes to animation style reproduction!
thanks man!
I've been hoping for this model since I arrived on SD!
Now all we need is more characters from Gargoyles (especially the milfs like Captain Chavez, Princess Katherine, Fox Xanatos... I can't do LoRA yet so I tried the hard way...it was a miserable failur that kept me awake the whole night!
By the way, any plan to bring this model or DCAU over to the SDXL side (of the Force), I've seen (and used the Arcane Lora XL over there and the result blew my mind. Seriously, when I saw the render A1111 did of Cassandra in casual sportswear I was blown away, and I though 'Can't wait till Lycon brings DC and Western over'.
VAE?
In my opinion, this is the best model on Civitai that conveys the flat style of American cartoons of the 70s-80s! Huge variety and flexibility. A cool model!
True! i hope is not discontinued for ever, i can't find anything even remotely similar
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.