Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Midlibrary: Library of 4000 Midjourney artistic styles

midlibrary.io

Post reply on HN
Print view View original

11–20 of 143 posts

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#11

Post by GaggiX »

brucethemoose2 wrote: post #10
GaggiX wrote: post #3

If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest , a list of 470 artists using the last model of NovelAI (this time they trained their model without removing artist names). Fun fact, this model is based on SDXL, finetuned probably on a cluster of H100s (from coreweave).

Lots of people (including potential trainers) struggle with SDXL I think. It just doesn't give you good results if you shove it into SD 1.5 frameworks and prompt it like SD 1.5.

If you're talking about anime, every popular SD v1.5 model out there has a backbone based on the leaked NovelAI model, so what SD XL lacks the most is people with compute power finetune it, that being said, this model: https://huggingface.co/cagliostrolab/animagine-xl-3.0 was recently released and it was finetuned for +500 A100-hours.

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#12

Post by Jackson__ »

brucethemoose2 wrote: post #10
GaggiX wrote: post #3

If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest , a list of 470 artists using the last model of NovelAI (this time they trained their model without removing artist names). Fun fact, this model is based on SDXL, finetuned probably on a cluster of H100s (from coreweave).

Lots of people (including potential trainers) struggle with SDXL I think. It just doesn't give you good results if you shove it into SD 1.5 frameworks and prompt it like SD 1.5.

The devs behind NovelAI have been talking about their training on the discord a little bit. According to their messages, it took about the same amount of compute to finetune the model, as it took for StabilityAI to train SDXL itself.

No wonder other people struggle to train it, if it takes >=128 A100s to get the model to perform better than SD1.5 on only slightly OOD tasks (anime pictures, in this case.)

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#13

Post by echelon »

Der_Einzige wrote: post #4
GaggiX wrote: post #3

If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest , a list of 470 artists using the last model of NovelAI (this time they trained their model without removing artist names). Fun fact, this model is based on SDXL, finetuned probably on a cluster of H100s (from coreweave).

The consensus seems to be that the lack of open source SDXL anime fine tunes is simply due to laziness and lack of resources from the interested parties in the community. Datasets like from the boorus are already floating around online to train on. Always funny to see that furries are ahead of weeaboos in the war for the best open source nsfw generator

> Always funny to see that furries are ahead of weeaboos in the war for the best open source nsfw generator

That's hilarious.

It'd be funny if they had a rivalry going and it resulted in better open source tooling.

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#14

Post by drewcoo »

nomilk wrote: post #2

Site no longer loading (anyone else?). Possible hug of death from hn traffic: > A timeout occurred. Error code 524. Visit cloudflare.com for more information. 2024-01-13 04:34:06 UTC Wayback machine gives a snapshot as at a few days ago: https://web.archive.org/web/20240113031449/https://midlibrar...

Mid-journey indeed!

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#15

Post by numpad0 »

Der_Einzige wrote: post #4
GaggiX wrote: post #3

If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest , a list of 470 artists using the last model of NovelAI (this time they trained their model without removing artist names). Fun fact, this model is based on SDXL, finetuned probably on a cluster of H100s (from coreweave).

The consensus seems to be that the lack of open source SDXL anime fine tunes is simply due to laziness and lack of resources from the interested parties in the community. Datasets like from the boorus are already floating around online to train on. Always funny to see that furries are ahead of weeaboos in the war for the best open source nsfw generator

I think that's just reflecting different levels of ethics standards - a bit like different stances on street graffiti

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#17

Post by thih9 »

What would happen if a platform focused on a company like Disney, offering images in the style of Frozen / Cinderella / etc?

Midjourney can already generate fire Elsa[1] and zombie Elsa[2]. I wonder if this kind of usage ever grows enough to start bothering corporations.

[1]: https://prompthero.com/prompt/be1b58e6fef

[2]: (nsfw cartoon gore) https://prompthero.com/prompt/00e7433fb1a

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#19

Post by nness »

imhoguy wrote: post #18

Dunno, I feel sad about artists a bit. I hope they won't stop creating. I am sick of these flashy AI fakes.

I guess at the end of the day, they're just fakes like any other fake reproduction. They'll be sold on marketplaces, occasionally a few buyers will be duped. But most are buying the style, not the artist, and wouldn't care.

The issue for this site is that they're using the artist names against the reproductions -- that could be argued as intentionally misrepresenting the origin of the works. I imagine, in no short order, they'll remove the artist's names or otherwise change the language to reflect that these are not the artists original works.

I'm not sure what people would use these for, other than their own personal use? Maybe inspiration, mood-boarding, or concept art... Some printing at home? I don't know.

Re: Midlibrary: Library of 4000 Midjourney artistic styles

#20

Post by ilaksh »

thih9 wrote: post #17

What would happen if a platform focused on a company like Disney, offering images in the style of Frozen / Cinderella / etc? Midjourney can already generate fire Elsa[1] and zombie Elsa[2]. I wonder if this kind of usage ever grows enough to start bothering corporations. [1]: https://prompthero.com/prompt/be1b58e6fef [2]: (nsfw cartoon gore) https://prompthero.com/prompt/00e7433fb1a

Yes, you can get Disney style and a lot of other copyrighted stuff. Yes, they are very bothered.
Post reply on HN