Chainclaw posted:

are there any AI projects that use midi files as input instead of actual audio?

Sedgr posted:

No tinkering these are unedited. Its interesting how changing the resolution changes the composition. Also, for anyone else using the Automatic1111, you can go into the ui-config JSON and change the Batch count max to whatever. I was running into issues because I couldnt do more than one image in a row changing the batch size without hitting a mem issue. But I could change the batch count and do up to 16. So now with the max changed I can do much larger batches.

TIP posted:

I was having the same issue and aside from the reboot I also made one change to the installation that might be what fixed it. During startup of SD I noticed a warning that said something like "LDSR is missing" and said that it was missing a "latent-diffusion" folder.

So I googled and found this:
https://github.com/CompVis/latent-diffusion

I downloaded that repo and put it in the spot specified in the error and changed the folder name to match (the repo's folder name was longer, shortened it to just latent-diffusion).

I didn't mention it earlier because upon restart I got a different error when it couldn't find another file in latent-diffusion. I couldn't figure out where to find that other file so I gave up on that and figured I had added the wrong repo, but I guess it might be what made it work.

AARD VARKMAN posted:

a geode split open to display a terrifying cosmic horror inside, elegant intricate digital painting artstation concept art by Mark Brooks and Brad Kunkle detailed
Steps: 50, Sampler: Euler a, CFG scale: 7.5, Size: 512x640

Snowy posted:

A bagpiper split open to display a terrifying cosmic horror inside, elegant intricate digital painting artstation concept art by Mark Brooks and Brad Kunkle detailed

AARD VARKMAN posted:

TheWorldsaStage posted:

Sedgr posted:

VectorSigma posted:

Neat!

BARONS CYBER SKULL posted:

roast chicken top down food photography 4k canon 80mm

tiling mode

Sedgr posted:

Those last ones were all "colorful painting, ((cyberpunk)), Quetzalcoatl, dark, gothic, detailed, intricate, geometric"

Most everything Ive posted was a variant of that replacing the god.

Rutibex posted:

try janus, he has two heads

TheWorldsaStage posted:

As sad as it is, Looney Tunes fell the gently caress off. I remember even adults wearing LT merchandise in the 90's, but when Cartoon Network stopped playing those kind of things if fell off pop culture's radar, so the AI not knowing Bugs or Tweety etc makes sense.

WhiteHowler posted:

Euler-a will exactly reproduce a previous prompt's output as long as you keep all input values the same, including the number of sampling steps. I don't exactly understand the logic, but I believe the "-a" variants use previous steps' output to inform the next step, so modifying the number of steps by even one can produce a drastically different result.

Other sampling modes tend to converge on an extremely similar result (both with themselves and with each other) above a certain threshold of sampling steps. This is why if you use the same seed for DDIM, the output at 40 steps should look almost identical to the output at 50 steps. Which should look very similar to the output from LMS at 40/50 steps.

If you do the same with Euler-a or DPM2-a, you'll usually get drastically different results at 40 and 50 steps.

There are several examples of this on Reddit. This post is good -- it shows the same prompt/seed/CFG values across the different sampling methods, and their output at different step counts. Note how Euler-a and DPM2-a have completely different results, both from the other methods and from themselves at different step counts.

TIP posted:

Oh drat, this just made me realize Euler-a and DPM2-a are incredibly valuable in that you can generate a bunch of variations at low computational cost.

Is there currently any way to automatically generate a certain number of steps and save a certain range of them as images? It seems like running a single seed with 50 steps and saving 20-50 would be an incredible way to get 30 variations on a prompt very quickly.

Elotana posted:

I know this is the art thread but I figure someone here knows: what's the current state of play for writing AI? The OpenAI content filter isn't as punishing on GPT-3 as it is on D-2 but I've tripped it a few times and I'd rather not lose access to both figuring out where the line is. Is there a cheaper quality service or an SD equivalent that I can run offline (with hopefully less GPU requirements and possibly a GUI of some kind?)

I'm not looking to generate college essays or SEO summaries; my use case is cranking the temperature way up for something like dialogue or short sentences, re-running it a bunch, then panning the reams of incoherent gibberish for the nuggets of humor/creativity.

Wheany posted:

Follow these instructions to get started: https://huggingface.co/docs/transformers/model_doc/gpt_neo

edit:

With 8GB of VRAM, I can also use the 2.7 billion parameter model by using half precision:

Python code:

from transformers import GPTNeoForCausalLM, GPT2Tokenizer
import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"

model = GPTNeoForCausalLM.from_pretrained("EleutherAI/gpt-neo-2.7B").half().to(device)
tokenizer = GPT2Tokenizer.from_pretrained("EleutherAI/gpt-neo-2.7B")

prompt="butts, lol"

input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)

num_new_tokens = 50
temperature = 0.95
TOP_P = 0.9
repetition_penalty = 1.5

gen_tokens = model.generate(
                input_ids,
                do_sample=True,
                temperature=temperature,
                repetition_penalty=repetition_penalty,
                top_p=TOP_P,
                max_new_tokens=num_new_tokens,
                pad_token_id=tokenizer.eos_token_id
            )

print(tokenizer.batch_decode(gen_tokens)[0])

e2: it's not quite GPT-3, but it's pretty good

Boba Pearl posted:

if you're using automatic-1111 yeah.

Wheany posted:

edit: This requires a lot of ram. I have 32GB on my computer and if i'm using more than ~12GB, the computer will swap like crazy while it's loading the model. Once it gets to the generating part ("gen_tokens=...."), the ram usage will plummet. But I've had my computer become completely unusable, mouse freezing and clock not updating for almost an hour when I tried to run this when I didn't have enough ram free.

Wheany posted:

No, it's Stable Diffusion specifically not knowing who Bugs Bunny is. Craiyon knows who he is and is doing its best:

Swedish Thaumocracy posted:

The Thinker statue as a rat. In a museum.


Realistic wild European dragon, in a mountain setting. highly detailed, ultra realistic, 8k resolution wildlife photography by Sir David Attenborough


gigantic muscly brown elephant dragon hybrid monster rampaging over iranian wilderness. Scaly skin. Very scary. Demonic face. [There were a few more prompt words here but automatic1111 didn't save them and I forgot what they were]






And finally, some star-dressed women! Bonus points for guessing the inspiration.

Beautiful Pink haired young woman wearing a purple elegant mature gorgeous pure dress made out of a sea of stars. Purple Victorian dress.


Beautiful ((Pink haired)) young asian woman, thin face, large yes, wearing a purple elegant mature gorgeous pure dress made out of a sea of stars.


A beautiful young asian woman with long straight (pastel pink hair). large pretty eyes. smiling. Wearing a (((purple gauze ball gown))) decorated with stylized golden LED stars.


All with Euler A, varying CFG and sample sizes of 18-30ish.

KinkyJohn posted:

How is the progress on ai voiceover generation? I want to have Morgan Freeman narrate some stupid videos

mobby_6kl posted:

Welp that's not happening with 12 gigs of total RAM lol.

Sedgr posted:

Tentacle Horrors.