AI Waifus and You 101

[Todd's Strongest Howard]

Figured I'd throw this here. Been fucking around with an online generator for meme purposes and despite my prompt, when doing basically any anime style from the dropdown menu it keeps spitting out images with this girl in it in most cases instead of Todd Howard.



I've been at this for an hour and it just keeps getting better and somehow Todd Howard is the key to this creation. At this point I don't think it's Todd's Mistake anymore, this is clearly Todd's Intent.

Spoiler: Jesus help me

Is Todd Howard the key to perfect waifus?

 

[RestlessRain]

Is Todd Howard the key to perfect waifus?

Considering how ugly the women are in any Bethesda game that isn't heavily modded... no, no he is not.

 

[Clem the Gem]

Not Vtuber related, but apparently, there's been a huge breakthrough in text to video generation. By Chrome's OpenAI team, no surprise there.

Sora: Creating video from text

openai.com

It's pretty baffling. There are still telltale signs that it's AI, but at the same time it manages to take into account crazy details like reflectivity during light scenes vs dark ones like this one in particular:



Like what...

All of the ones featured on their article are truly insane to me.

Figured I'd throw this here. Been fucking around with an online generator for meme purposes and despite my prompt, when doing basically any anime style from the dropdown menu it keeps spitting out images with this girl in it in most cases instead of Todd Howard.

View attachment 64521

I've been at this for an hour and it just keeps getting better and somehow Todd Howard is the key to this creation. At this point I don't think it's Todd's Mistake anymore, this is clearly Todd's Intent.

Spoiler: Jesus help me

View attachment 64525View attachment 64527View attachment 64528View attachment 64529View attachment 64530

Is Todd Howard the key to perfect waifus?

If you really are not prompting anything else, my guess would be either "todd" or "howard" can be found in the name of a character in some anime and that's where it's pulling it from. I did a quick search but didn't find anything though.
A nice happy accident anyawy.

[Edit] Quoted wrong post..

 

[Koronesuki]

You've created genderswapped anime Todd Howard. Why would you do this?

 

[Faceless Waifu]

You've created genderswapped anime Todd Howard. Why would you do this?

For the glory of satan, of course!

 

[Todd's Strongest Howard]

All of the ones featured on their article are truly insane to me.



If you really are not prompting anything else, my guess would be either "todd" or "howard" can be found in the name of a character in some anime and that's where it's pulling it from. I did a quick search but didn't find anything though.
A nice happy accident anyawy.

[Edit] Quoted wrong post..

I think for the most part it's a weird collision of the keywords firing off something crazy. The base prompt I used that I worked off of is "Todd Howard, facing the camera, with the body of a woman, wearing a low cut shirt and dark brown skin, with a halo of golden light radiating from his head". A lot of the time it doesn't seem to really process through well and more often than not it gives me Super Todd or various women, but I realize that broad strokes visual elements of Todd are carried through, notably the hair. The rest of the elements kind of pick up from there for a consistent look.



Both of these were generated off the same prompt, on the same page. It was looking at the hair that made me realize that it's basically scanning the general likeness of Todd and feminizing it while not delivering on his likeness. I was trying to make a better image to use for an avatar to hop on the retard trend of brown tiddy pfps someone mentioned somewhere with a Todd flair but I accidentally ended up with brown women. Truly a shame.

 

[Watamate]

I think for the most part it's a weird collision of the keywords firing off something crazy. The base prompt I used that I worked off of is "Todd Howard, facing the camera, with the body of a woman, wearing a low cut shirt and dark brown skin, with a halo of golden light radiating from his head". A lot of the time it doesn't seem to really process through well and more often than not it gives me Super Todd or various women, but I realize that broad strokes visual elements of Todd are carried through, notably the hair. The rest of the elements kind of pick up from there for a consistent look.

View attachment 64538View attachment 64539

Both of these were generated off the same prompt, on the same page. It was looking at the hair that made me realize that it's basically scanning the general likeness of Todd and feminizing it while not delivering on his likeness. I was trying to make a better image to use for an avatar to hop on the retard trend of brown tiddy pfps someone mentioned somewhere with a Todd flair but I accidentally ended up with brown women. Truly a shame.

You could try adding the tags "1guy, solo", into your prompt and maybe "1girl" into the negative prompt. My personal guess is whatever model it's using is heavily trained on female character/images. I also don't know if the web generator you're using can use Loras (if that's even a thing with online generators), if it can however you can try using this one:

Todd Howard - 1 | Stable Diffusion LoRA | Civitai

This model has been trained on images of the one and only Todd Howard. Works well with Deliberate with 1.0 strength.

civitai.com

If you want something specific, I can also just try making something myself if you give me enough info to go on.

 

[Todd's Strongest Howard]

You could try adding the tags "1guy, solo", into your prompt and maybe "1girl" into the negative prompt. My personal guess is whatever model it's using is heavily trained on female character/images. I also don't know if the web generator you're using can use Loras (if that's even a thing with online generators), if it can however you can try using this one:

Todd Howard - 1 | Stable Diffusion LoRA | Civitai

This model has been trained on images of the one and only Todd Howard. Works well with Deliberate with 1.0 strength.

civitai.com

If you want something specific, I can also just try making something myself if you give me enough info to go on.

It's perchance, and this is just an oddity of the anime filters. There's a dropdown menu with a bunch of other pre-designed prompt packs that turn out interesting results, but I was rolling through looking for various styles and/or future pfp replacements when I hit Anime and it started spitting waifus instead of Howies. This is pretty consistent for any of the anime styles, especially waifu.

Besides, I'm not looking for "good", I'm looking for "shitpost". The uncanny nature of a lot of Todd generations is really funny regardless, and a shocking amount of the time it just comes out looking like an old lesbian.

e: Man I'll tell you what, usually text doesn't pop up that clearly in AI images but in this case I'm shook.



       TODDIM

 

[Faceless Waifu]

I got bored. Here's what I made using Tensor.Art, a free-ish StableDiffusion-style AI Art generator that is a bit less "free" compared to Civitai or PixAi, but is arguably has cheaper access to much-needed HiresFix for that good looking results (PixAI cost too big for constant generations with HiresFix and locked you to one image per generation, while Civitai at the moment has no means to use HiresFix at all).

Selen Tatsuki, formerly of Nijisanji EN. AI Generated using Tensor.Art using Kaoming (Dev_Is Raden's Mama) artstyle LORA and a Selen Tatsuki LORA.

Revine Millith, a 1view vtuber that pops out thanks to The Proctor's 'managerial skills' back during the CyberIdol shenanigans. AI Generated using Tensor.Art using Kaoming LORA and prompted using what I can guess from the only existing art of hers (so it's not 100% similar to her avatar. Had to make do.)

 

[Koronesuki]

Those are nice, although I notice it didn't get Selen's heterochromia right. (Also the hands, AI can't ever seem to get hands right.)

 

[Faceless Waifu]

Those are nice, although I notice it didn't get Selen's heterochromia right. (Also the hands, AI can't ever seem to get hands right.)

Selen's heterochromia one was me forgetting to prompt it. And even with better hands LORA, the curse of AI hands still strikes. I might try doing more selen once my daily credits is reset, or try to see if Tensor.Art has dokibird LORA lurking around.

For now, have this one that I generate using Chainsawman Anime(MAPPA Studios) style LORA, which couldn't able to recreate Selen's head-wings but still looks fancy enough not to share. I did try generating it with head-wings but the result wasn't satisfying enough (and it gave her wings on her back, which is not what I want).

 

[Koronesuki]

Personally, I think she looks better without the head-wings, but I realize the goal here is to generate results that actually look like the character, so there's still room for improvement.

 

[Watamate]

Selen's heterochromia one was me forgetting to prompt it. And even with better hands LORA, the curse of AI hands still strikes. I might try doing more selen once my daily credits is reset, or try to see if Tensor.Art has dokibird LORA lurking around.

For now, have this one that I generate using Chainsawman Anime(MAPPA Studios) style LORA, which couldn't able to recreate Selen's head-wings but still looks fancy enough not to share. I did try generating it with head-wings but the result wasn't satisfying enough (and it gave her wings on her back, which is not what I want).

Best bet with hands and stuff is using ControlNet and OpenPose. But even that is honestly a crapshoot, but definitely better than nothing. That's if the site you're using supports that though.

 

[Alchemyth]

https://arxiv.org/html/2402.17764v1

Potentially revolutionary architecture. Relevant to your waifus since SD is transformers based. If this pans out, they could train a massive model, for a quarter of the cost and performance footprint.

 

[Clem the Gem]

https://arxiv.org/html/2402.17764v1

Potentially revolutionary architecture. Relevant to your waifus since SD is transformers based. If this pans out, they could train a massive model, for a quarter of the cost and performance footprint.

That's a lot of words and numbers. I'm sure it's good news, but right now I'll just have to take your word for it.

 

[Alchemyth]

That's a lot of words and numbers. I'm sure it's good news, but right now I'll just have to take your word for it.

For LLM's, a 70b parameter model at full precision requires roughly 140GB of ram to load the entire model. Under this new architecture, it would take only 18GB. It's insane.

 

[God's Strongest Dragoon]

Also the hands, AI can't ever seem to get hands right.

True, but neither can a lot of twitter artists that sperg about AI drawing hands

 

[Koronesuki]

True, but neither can a lot of twitter artists that sperg about AI drawing hands

A lot of Twitter users probably couldn't pass a Turing test either.

 

[Watamate]

I'm probably in the minority when it comes to using Chatbots/SillyTavern. Though, if people are interested, they can still reach out to me. There's a fairly good free proxy available currently.

But to get back on topic, Anthrophic just released Claude 3. Claude being the main competitor to GPT and I honestly prefer it for most cases. I will say that Gemini has made leaps and bounds with their more recent models giving both of them a run for their money. Now I haven't tried it out yet, and since it only became available today I'm not sure if it'll work with SillyTavern yet (maybe via the Custom Chat Completion UI). But I've got to say it sounds incredibly promising.

Introducing the next generation of Claude

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

www.anthropic.com

 

[Watamate]

So I stumbled on a OC character I liked... But for some fucking reason the person seemed to ruin it by making it a futa. So I tried (I think successfully) to recreate it. Obviously, I'm not going to post the original picture. But I do wonder, is there any vtuber out there with a similar vibe? Not really NSFW, but still.



It was also my first time using an SDXL-based checkpoint, and I do quite like it so far. Besides the fact that none of my old Loras are compatible and I have to start from scratch... But that's fine right? ...