Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
the impression I got from it was Kurzweil is so desperate to live forever that there's no line he's unwilling to cross to get there. If he could make himself a head in a jar like Futurama, he absolutely would. He seemed to think that were he born 10 years later, he might've been able to see a world of vastly extended lifespans, so he was desperately micromanaging his diet and exercise to get himself over the line.

I was a teenager when I watched it and looked forward to this idea of people living 100s of years, thinking it was very cool. One consequence of getting older that nobody warned me about was the loss of hopeful naïveté since these ideas now inspire dread and fear within me.
It will take a true gigachad to embrace immortality. Dying is for loser cavemen.

Exponential growth brews 1 million AI models on Hugging Face (archive)
 
It's also hampered by the laughably small 8K context limit. They've based Erato on Llama 3 whose initial release had a gimpy context limit, but 3.1 can handle much more. If they retrained on 3.1, they could potentially stretch to 32K context limits, though I don't see this happening anytime soon as they've likely invested millions of dollars into the training they've done with 3 and are unwilling to start over.
If you are interested in long context sizes Mistral Nemo is the good shit. 128K context size and it's surprisingly good for a 12B model.
 
  • Like
Reactions: Vecr and Riftrunner
Are there any good free in-browser text AIs?

I used to use the OpenAI text playground, even bought some credits when the free beta ended, but it became a bit steep for the results it started to spit out. Perhaps my use of it was too niche? I host an RP map game (I know, I eagerly await the puzzle pieces) and would use stuff that I or players had written as prompts and watch it fly essentially. It was good for a laugh (like players suddenly turning on friends and allying with AI halluciNations) and for spinning off ideas, and it'd be nice to do again for my newest batch of players.
 
They do normal text completion, not instruct, yes? Llama 3.0 (which I assume since 3.1 wasn't explicitly named and 3.2 literally just came out and isn't different in the text department anyways) has looping issues that are especially noticeable in the base model, I wonder if a literature-centric finetune fixed them. The architecture of the base llama 3.0 models is somehow not really useful for longer contexts (4096+)they break down incredibly quickly. I think they were trained on very short chunks of material and even with finetunes with current gen dense models, it seems impossible to overcome that. I can't really imagine a model suited worse for text completion in a scenario like this tbh
 
They do normal text completion, not instruct, yes? Llama 3.0 (which I assume since 3.1 wasn't explicitly named and 3.2 literally just came out and isn't different in the text department anyways) has looping issues that are especially noticeable in the base model, I wonder if a literature-centric finetune fixed them. The architecture of the base llama 3.0 models is somehow not really useful for longer contexts (4096+)they break down incredibly quickly. I think they were trained on very short chunks of material and even with finetunes with current gen dense models, it seems impossible to overcome that. I can't really imagine a model suited worse for text completion in a scenario like this tbh
There are some work arounds with ROPE in illama.cpp for repetition. Really though fine-tunes along with context injection from another data source are the only way to make it useful. All that silly tavern is is just context injection based on scanning for keywords. Plus auto summarizing data as it is being dropped from the context.

I think the next advancement in LLMs isn't going to be the LLM, but how to efficiently manage context via algorithms. Basically what needs to happen is for a frontend to be able to detect uneeded but important information and stash it somewhere while removing it from the context For example:

"On the horizon the Kiwi's domain stands. The castle is on a hill overlooking a scenic field where many cows graze. Jack approaches on his steed."

Would be tokenized to "Jack approached Kiwi's domain, and later left", as the story progresses

With the description of Kiwi's domain injected into the context as only when the story requires a complex description of it.

Instruct: I go back to Kiwi's Domain.

Context-Injection: Kiwi's domain is a castle standing on a hill overlooking a field full of cows.

Response:Jack approached the castle once again weary from his journey. He felt soothed looking at the field of docile animals from the hill on which the castle stood.
 
News hams (and special guest, Buttah Dawg)
Hammy Halloween!!!
0-1.jpg0-2.jpg
1-1.jpg
1-2.jpg1-3.jpg1-4.jpg1-5.jpg1-6.jpg1-7.jpg
2-2.jpg
2-3.jpg2-5.jpg2-4.jpg
3-1.jpg
3-2.jpg3-3.jpg
4-1.jpg
4-2.jpg4-3.jpg4-4.jpg4-5.jpg
edit to add after catching up on MatI:
5-1.jpg
5-2.jpg5-3.jpg5-4.jpg5-5.jpg5-6.jpg
 
Last edited:
Meta AI occasionally makes gorillas with the prompt black guy. It also has cool editing features. I think it's in messenger only but this is like Xbox Kinect racism

Screenshot_20241001-091716.pngScreenshot_20240927-125151.pngScreenshot_20240927-125352.png

Realized this isn't a proper display it also makes noncrap -
Img_2024_10_02_00_43_49.jpeg
Img_2024_10_02_00_45_28.jpeg
 

Attachments

  • Screenshot_20240930-223050.png
    Screenshot_20240930-223050.png
    674.6 KB · Views: 28
Last edited:
Meta’s new “Movie Gen” AI system can deepfake video from a single photo (archive)

This is yet another unreleased thing, but it can give you an idea of where things are headed. It's not just animating a photo, but creating unrelated video from a photo and text prompt:


A Movie Gen deepfake video with the prompt: "Make a cute selfie video of a man and his dog. The man is wearing a black shirt. The dog is a beagle puppy. The background is a backyard patio, filled with trees. The man has a big smile on his face, as he tries to take the perfect selfie with his dog. The lighting is warm." The original source photo is in the upper-right corner.

It can also generate videos from nothing but text as usual, edit existing videos, and add sound effects and music using the prompt.

 
  • Thunk-Provoking
Reactions: anustart76
Is there a good general llm and platform for Linux? I just need to do shit like make text more concise.
 
Is there a good general llm and platform for Linux? I just need to do shit like make text more concise.
Ollama if you just want to run a model without any extra work, they have a bunch of models available in their library you can just pull directly. For just summary Llama 3.2 8b is good enough unless you run into censorship issues, in which case Dolphin-Mixtral is pretty good, I would try out a bunch to see which fits your usecase best. They also provide docs on how to use LangChain to vectorize a large document to query here

If you want to do more than that I would recommend checking out KoboldCPP which allows a lot more control and a nice web interface for you to use. Both Ollama and Kobold are frontends for Llama.CPP

There is also Sillytavern which was made for gooners, but it also works as a decent frontend with all the settings exposed and character cards for you to define instructions. It also allows you to hook up multiple APIs all at once so you can have everything in one place. But I still think for specific tasks Ollama is probably the better option.
 
Last edited:
  • Like
Reactions: Palmer Bangs
Is there a good general llm and platform for Linux? I just need to do shit like make text more concise.
For the platform I'd start with: https://github.com/oobabooga/text-generation-webui
I have no recommendations for actual models though.

I tried to get them to summarize some of your MATI streams and there was just too much length, although it's been a while so maybe things have improved.

A lot of this shit is still shit, or a bunch of disjointed Python you have to assemble yourself.

For instance: https://huggingface.co/facebook/bart-large-cnn
 
finding a model that isn't made exclusively for coomers is the biggest challenge though. lol

on the bright side you can make your own character file to use with the models, so it can help null make that Chantal sex bot he's always wanted.
I was able to get whatever LLM I was using for the MATI stuff to answer questions about Josh. Including important things like "What is Josh's favorite cheese?" the answer, of course, was cheddar.

I'm thinking this AI stuff has a promising future.
 
  • Informative
Reactions: Train Operator
Ollama if you just want to run a model without any extra work, they have a bunch of models available in their library you can just pull directly. For just summary Llama 3.2 8b is good enough unless you run into censorship issues, in which case Dolphin-Mixtral is pretty good, I would try out a bunch to see which fits your usecase best. They also provide docs on how to use LangChain to vectorize a large document to query here

If you want to do more than that I would recommend checking out KoboldCPP which allows a lot more control and a nice web interface for you to use. Both Ollama and Kobold are frontends for Llama.CPP

There is also Sillytavern which was made for gooners, but it also works as a decent frontend with all the settings exposed and character cards for you to define instructions. It also allows you to hook up multiple APIs all at once so you can have everything in one place. But I still think for specific tasks Ollama is probably the better option.
Sillytavern has some interesting features. World Books or something like them are the future of LLM application. They allow you to introduce custom concepts or data into the LLM by defining those concepts along with a keyword trigger. For example you could inject a user's account balance or some other setting which the LLM could use. This eliminates the need for trivial fine-tunings.
 
Is there a good general llm and platform for Linux? I just need to do shit like make text more concise.

For the platform I'd start with: https://github.com/oobabooga/text-generation-webui
I have no recommendations for actual models though.

I tried to get them to summarize some of your MATI streams and there was just too much length, although it's been a while so maybe things have improved.

A lot of this shit is still shit, or a bunch of disjointed Python you have to assemble yourself.

For instance: https://huggingface.co/facebook/bart-large-cnn

finding a model that isn't made exclusively for coomers is the biggest challenge though. lol

on the bright side you can make your own character file to use with the models, so it can help null make that Chantal sex bot he's always wanted.

Right now, I like Mistral-Nemo. Pretty good general purpose model.
 
Is there a good general llm and platform for Linux? I just need to do shit like make text more concise.
Llama is a rare Facebook W, the 8B variant of 3.1 can run on consumer hardware well
Use ollama to manage it and it will work as a REST API with tokenization built in
 
Back