AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

EVIL Gibson: Mar 23, 2001; Internet of Things is just someone else's computer that people can't help attaching cameras and door locks to!; Switchblade Switcharoo

LordAdakos posted:

I'm having the same issues. Any SwinIR model I try gives me a pytorch size error or key Error for params_ema.

[hahahaha pytorch being pytorch]

EDIT: I posted a third party tool for Anaconda at the end of this gripe fest that can sometimes help with the requirement problem.

Man, you are all now realizing how lovely pytorch is which I have been frustrated with for the past four years.

That loving module just decides with a dart board which functions are removed and added as it gets new versions.

"Oh, but EVIL Gibson you can just install a lower version of the module to get it to work! Just use the requirements installer"

Heres the absolute hosed part. That requirements file might have worked when it first came out, but the modules either no longer support the versions of CUDA available or the math libraries just changed and there is no way to all the other requirements to be supported at the same time!

You literally have to break your python environment to get anything to work especially now when the developers have not tested their current environment forgetting they bodged/forced a module to be loaded when it technically shouldn't work... and forget to include that module.

In summary, all of you please complain directly to the devs/users about this poo poo that has been going on for a long time

Now the tip:
To help with requirements, I found a replacement for "conda" (for Anaconda) called "mamba".

quote:

mamba is a reimplementation of the conda package manager in C++.

parallel downloading of repository data and package files using multi-threading
libsolv for much faster dependency solving, a state of the art library used in the RPM package manager of Red Hat, Fedora and OpenSUSE
core parts of mamba are implemented in C++ for maximum efficiency
At the same time, mamba utilizes the same command line parser, package installation and deinstallation code and transaction verification routines as conda to stay as compatible as possible.

basically, think of it as a faster and waaaay smarter requirement solver for "conda". You will find it able to figure out what an application needs that conda will say "Haha, that is impossible and you are a fool for trying to install this :derp:

"

https://github.com/mamba-org/mamba

EVIL Gibson fucked around with this message at 22:43 on Sep 23, 2022

# ? Sep 23, 2022 22:30

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 02:42

mobby_6kl: Aug 9, 2009; by Fluffdaddy

In fact all of python can go die in a fire

# ? Sep 23, 2022 22:31

Rahu: Feb 14, 2009; let me just check my figures real quick here; Grimey Drawer

EVIL Gibson posted:

Man, you are all now realizing how lovely pytorch is which I have been frustrated with for the past four years.

That loving module just decides with a dart board which functions are removed and added as it gets new versions.

"Oh, but EVIL Gibson you can just install a lower version of the module to get it to work! Just use the requirements installer"

Heres the absolute hosed part. That requirements file might have worked when it first came out, but the modules either no longer support the versions of CUDA available or the math libraries just changed and there is no way to all the other requirements to be supported at the same time!

You literally have to break your python environment to get anything to work especially now when the developers have not tested their current environment forgetting they bodged/forced a module to be loaded when it technically shouldn't work... and forget to include that module.

In summary, all of you please complain directly to the devs/users about this poo poo that has been going on for a long time

Always grab a package list with versions once you have it working. Dependency management is the devil.

# ? Sep 23, 2022 22:40

EVIL Gibson: Mar 23, 2001; Internet of Things is just someone else's computer that people can't help attaching cameras and door locks to!; Switchblade Switcharoo

Rahu posted:

Always grab a package list with versions once you have it working. Dependency management is the devil.

I posted the best tool for dependency management in that same post so the entire post was not 100% bitching

https://github.com/mamba-org/mamba

It solves dependency issues like mad that conda would have given up on. You can use it as an entire replacement to conda and it just adds many more features like repoquery

code:


$ mamba repoquery depends -t xtensor

xtensor == 0.21.5
&#9500;&#9472; libgcc-ng [>=7.3.0]
&#9474; &#9500;&#9472; _libgcc_mutex [0.1 conda_forge]
&#9474; &#9492;&#9472; _openmp_mutex [>=4.5]
&#9474;   &#9500;&#9472; _libgcc_mutex already visited
&#9474;   &#9492;&#9472; libgomp [>=7.3.0]
&#9474;     &#9492;&#9472; _libgcc_mutex already visited
&#9500;&#9472; libstdcxx-ng [>=7.3.0]
&#9492;&#9472; xtl [>=0.6.9,<0.7]
    &#9500;&#9472; libgcc-ng already visited
    &#9492;&#9472; libstdcxx-ng already visited

code:


$ mamba repoquery whoneeds ipython

Name            Version Build          Depends          Channel
-------------------------------------------------------------------
jupyter_console 6.4.3   pyhd3eb1b0_0   ipython          pkgs/main
ipykernel       6.9.1   py39haa95532_0 ipython >=7.23.1 pkgs/main
ipywidgets      7.6.5   pyhd3eb1b0_1   ipython >=4.0.0  pkgs/main

the latter example is absolutely sweet since it tells you the exact version the module needs so you can upgrade to the minimum required which would be the greatest version specific in (it being ipython v7.23.1 for ipykernel in the example)

EVIL Gibson fucked around with this message at 22:50 on Sep 23, 2022

# ? Sep 23, 2022 22:45

WhiteHowler: Apr 3, 2001; I'M HUGE!

Dear Stable Diffusion,

Please show me "bofa".

That'll do.

# ? Sep 23, 2022 22:47

TIP: Mar 21, 2006; Your move, creep.

someone posted about cornish rex cats, which are very skinny hairless cats, and I decided to combine them with rex goliath chickens (the largest chicken breed)

my prompt was "Cornish Rex Goliath chicken, photo, detailed, 4k"

the results may be :nms:

# ? Sep 24, 2022 00:52

Sedgr: Sep 16, 2007; Neat!

They feel like a tongue but taste like chicken!

# ? Sep 24, 2022 00:55

Comfy Fleece Sweater: Apr 2, 2013; You see, but you do not observe.

BARONS CYBER SKULL posted:

dramatic lighting, bust shot, (beautiful female vampire) skeletal robe smoke luminescent haunted ((((JPG glitch heads_up _display)))) evil inky swirly ripples, ((fluffy hair)) long_hair, sickly colors, (character sheet),detailed pupils, fine details (eyeliner) concept design concept art contrast kim jung gi greg rutkowski trending on artstation 8k, realistic proportions artgerm alphonse mucha

Negative prompt: ugly, fat, obese, chubby, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, penis, nose, eyes, lips, eyelashes, text,

Steps: 100, Sampler: DDIM, CFG scale: 8, Seed: 1925338945, Size: 512x704

What exactly do the parentheses do? emphasize the tag?

# ? Sep 24, 2022 02:14

Boba Pearl: Dec 27, 2019; by Athanatos

Comfy Fleece Sweater posted:

What exactly do the parentheses do? emphasize the tag?

Multiplies the weight of the prompt by 1.1 per set, which only works in the automatic-1111 build.

# ? Sep 24, 2022 02:15

Moongrave: Jun 19, 2004; Finally Living Rent Free

certain things REALLY need the boost or they get lost in the noise

# ? Sep 24, 2022 02:20

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

((baby seal)) disguised as a loaf of bread engages the enemy faction and fires the graviton gun from Syndicate Wars while at club Mar-a-Lago view from a news truck, Paul Lehr, Jeff Easley

# ? Sep 24, 2022 02:29

Rutibex: Sep 9, 2001; by Fluffdaddy

TIP posted:

someone posted about cornish rex cats, which are very skinny hairless cats, and I decided to combine them with rex goliath chickens (the largest chicken breed)

my prompt was "Cornish Rex Goliath chicken, photo, detailed, 4k"

the results may be

This is an excellent basilisk

# ? Sep 24, 2022 02:36

Sedgr: Sep 16, 2007; Neat!

Master of the petrifying gaze. Sentient silica. In vitro veritas.

# ? Sep 24, 2022 02:52

Comfy Fleece Sweater: Apr 2, 2013; You see, but you do not observe.

BARONS CYBER SKULL posted:

dramatic lighting, bust shot, (beautiful female vampire) skeletal robe smoke luminescent haunted ((((JPG glitch heads_up _display)))) evil inky swirly ripples, ((fluffy hair)) long_hair, sickly colors, (character sheet),detailed pupils, fine details (eyeliner) concept design concept art contrast kim jung gi greg rutkowski trending on artstation 8k, realistic proportions artgerm alphonse mucha

Stole some of your prompt and it pumped out some real good stuff :catstare:

Prompt was a bear, dramatic lighting, bust shot, fluffy hair, happy colors, fine details concept design concept art contrast kim jung gi greg rutkowski trending on artstation 8k, realistic proportions artgerm alphonse mucha

This one I thought was the best, it's ridiculously good imo. Whose style is this ?

Didn't give a gently caress about the "bear" being male, female, object, animal, etc, but that was a happy accident

DETECTIVE BEAR

# ? Sep 24, 2022 03:03

Snackmar: Feb 23, 2005; I'M PROGRAMMED TO LOVE THIS CHOCOLATY CAKE... MY CIRCUITS LIGHT UP FOR THAT FUDGY ICING.

Swedish Thaumocracy posted:

Tilda Swinton as Commander Shepard punching a reporter, movie still from the Mass Effect movie. Space sci-fi action video game movie still.

Oh I love this - I modified it to "Tilda Swinton as Cloud Strife punching a chocobo, movie still from the Final Fantasy 7 movie. Space sci-fi action video game movie." still and got some neat results (the ones with chocobos weren't great)

Snackmar fucked around with this message at 03:15 on Sep 24, 2022

# ? Sep 24, 2022 03:11

Comfy Fleece Sweater: Apr 2, 2013; You see, but you do not observe.

The master himself

# ? Sep 24, 2022 04:22

Tunicate: May 15, 2012

btw if you're in teh dalle2 discord, you can sign up for their october contest and get 5 free prompts per day for the month

# ? Sep 24, 2022 04:55

VectorSigma: Jan 20, 2004; Transform
and
Freak Out

just lovin the high res fix and LDSR

glowing gold river in a jardin extraterrestre constructed from chrome on the surface of (cybertron) by greg rutkowski makoto shinkai takashi takeuchi
Steps: 100, Sampler: Euler a, CFG scale: 30, Seed: 3606310980, Size: 2048x1024, Denoising strength: 0.7

# ? Sep 24, 2022 05:31

tote up a bags: Jun 8, 2006; die stoats die

VectorSigma posted:

just lovin the high res fix and LDSR

glowing gold river in a jardin extraterrestre constructed from chrome on the surface of (cybertron) by greg rutkowski makoto shinkai takashi takeuchi
Steps: 100, Sampler: Euler a, CFG scale: 30, Seed: 3606310980, Size: 2048x1024, Denoising strength: 0.7

This owns.
Are you using High res fix to generate it at full size, then throwing it to LDSR after that? Or generating small then LDSR scaling up?

# ? Sep 24, 2022 07:54

mobby_6kl: Aug 9, 2009; by Fluffdaddy

pixaal posted:

The documentation for automatic 1111 says (increase_weight) [decrease_weight] apparently by 10% it doesn't say anything about feeding it weights manually. you can (((stack_them)))

I have no idea if you can include multiple words so I used underscore include multiple words between at your own experimentation.

Thanks, that seems to work correctly, including multiple words with spaces such as "[mona lisa] (((cave painting)))"

# ? Sep 24, 2022 09:32

Lord Stimperor: Jun 13, 2018; I'm a lovable meme.

Burning Rain posted:

I mean, I wouldn't mind if Open AI/Midjourney/etc had to pay out like 10% of income to (in-copyright) artists used in prompts proportionally to how often they're invoked. Imo, Rutkowski should be compensated for inadvertently turning out the secret ingredient of fantasy AI art.

Just wait for my cool blockchain nft idea

# ? Sep 24, 2022 09:36

Lord Stimperor: Jun 13, 2018; I'm a lovable meme.

This post rocks imo. Watch a computer dream up scenes
https://www.reddit.com/r/StableDiffusion/comments/xm8nuw/i_tried_making_procedural_forest_animation_still/

# ? Sep 24, 2022 09:43

nightwisher: Dec 24, 2004

VectorSigma posted:

just lovin the high res fix and LDSR

glowing gold river in a jardin extraterrestre constructed from chrome on the surface of (cybertron) by greg rutkowski makoto shinkai takashi takeuchi
Steps: 100, Sampler: Euler a, CFG scale: 30, Seed: 3606310980, Size: 2048x1024, Denoising strength: 0.7

This looks so absolutely fantastic. Took your prompt and mucked around with it a bit, gave me this rather vibrant result, though the reflections are still not accurate :v:

Positives:((glowing gold river)) in a jardin extraterrestre constructed from (chrome) on the surface of (cybertron) by greg rutkowski makoto shinkai takashi takeuchi accurate_reflections panoramic night
Negatives: blurry, (inaccurate lighting), (boring)
Properties: Steps: 100, Sampler: Euler a, CFG scale: 25, Seed: 2997980240, Size: 2048x1024, Model hash: 7460a6fa, Denoising strength: 0.7
Upscaled with CupScale

# ? Sep 24, 2022 10:13

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

I change and mutate prompts a lot when they go through img2img to really say these have a prompt but here's what the original goal of each was.

Cute Slime eating cookie

Kitten Popping a balloon tied to her back

# ? Sep 24, 2022 13:33

WhiteHowler: Apr 3, 2001; I'M HUGE!

VectorSigma posted:

just lovin the high res fix and LDSR

glowing gold river in a jardin extraterrestre constructed from chrome on the surface of (cybertron) by greg rutkowski makoto shinkai takashi takeuchi
Steps: 100, Sampler: Euler a, CFG scale: 30, Seed: 3606310980, Size: 2048x1024, Denoising strength: 0.7

Would you please explain exactly how you're doing the larger pictures? Like, step by step? What hardware/VRAM do you have?

I'm running Automatic1111 with default settings on an RTX 2070 Super with 8 GB of VRAM. It does 512x512 just fine, but I get "out of VRAM" errors if I try going to 512x768. Obviously I can do a 512x512 and upscale to 1024x1024 or 2048x2048. I'm using the Extras tab and upscaling 2x or 4x, but I'd like to do some portrait or landscape stuff like people here are starting to do.

Do I just need a better GPU? I was waiting for the 40xx series, but Nvidia's pricing seems a little crazy next generation.

# ? Sep 24, 2022 13:59

Boba Pearl: Dec 27, 2019; by Athanatos

WhiteHowler posted:

Would you please explain exactly how you're doing the larger pictures? Like, step by step? What hardware/VRAM do you have?

I'm running Automatic1111 with default settings on an RTX 2070 Super with 8 GB of VRAM. It does 512x512 just fine, but I get "out of VRAM" errors if I try going to 512x768. Obviously I can do a 512x512 and upscale to 1024x1024 or 2048x2048. I'm using the Extras tab and upscaling 2x or 4x, but I'd like to do some portrait or landscape stuff like people here are starting to do.

Do I just need a better GPU? I was waiting for the 40xx series, but Nvidia's pricing seems a little crazy next generation.

Are you launching with med vram as your launch option? What about low vram?

# ? Sep 24, 2022 14:11

Comfy Fleece Sweater: Apr 2, 2013; You see, but you do not observe.

https://www.engadget.com/nvidia-ai-model-get3d-virtual-worlds-objects-130021127.html

3d is coming

https://www.youtube.com/watch?v=KfcsdhGKb1U

# ? Sep 24, 2022 14:18

WhiteHowler: Apr 3, 2001; I'M HUGE!

Boba Pearl posted:

Are you launching with med vram as your launch option? What about low vram?

Just using the default config. The only extra flag I'm giving it is "--listen" so my wife can generate her body horror images over the LAN.

# ? Sep 24, 2022 15:03

Wheany: Mar 17, 2006; Spinya^{ha^{ha^haha}ha}ha_{ha_{ha_haha}ha}ha!; Doctor Rope

I took a bunch of prompts from this thread and twitter (I edited them a little just so every prompt doesn't have "by Greg Rutkowski" in it). I let the computer take a random subset of them and then fed them into GPT Neo to generate more prompts:

"The original cat lady, art by Magali Villeneuve, digital painting, hd, hdr, 8k, digital painting, hdr, 8k, realism"

"Mountainous mountains, the Himalayas, the Alps, art by Franz Xaver Winterhalter, digital painting, high end, sharp focus, 8k, ray tracing, hdr, hd, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr, hdr"

"Norman Rockwell's paintings of birds nesting in old trees, oil painting, 1940s"

"The artist who made the most famous computer art piece ever is none other than David Ortiz, a baseball player for the Boston Red Sox, baseball card, picture frame, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card, baseball card"

"a woman looking through binoculars into a valley, photo by takafumi matsuoka, black and white"

"a baby chick peeking out from a nest, photo by michael smits, black and white"

"a horse drawing a bow, artwork by george bickmore, high definition, sharp focus, 8k, ray tracing, digital painting, concept art illustration by alphonse mucha, by guweiz and wlop"

"Pennywise the Dancing Clown from "It" film, movie still, 4k, blurry, blurry lens distortion, blurred eyes, blurried facial features, art by michael sommers, artstation"

"The world's smallest superhero, comic book illustrations, 8k, artstation"

"a man with a face that turns into the shape of a spider, digital painting, 8k, artstation"

# ? Sep 24, 2022 15:04

Vlaphor: Dec 18, 2005; Lipstick Apathy

More "gateway between life and death" goodness.

Sadly, I do not recall the full prompt. Been experimenting more with the Automatic1111 build. Before then, I was using HLKY, which is easier to use, but 1111 gives more options and fine tuning, which I'm still learning on.

# ? Sep 24, 2022 15:08

Lemon: May 22, 2003

Boba Pearl posted:

I will write you one my friend.

First Requirements for the thing I'm using:

Windows, Nvidia card

there's linux and amd distros, but I don't know how to use that.

Ok, go here:

https://github.com/AUTOMATIC1111/stable-diffusion-webui/

Unzip this to it's own folder. You download it by clicking Code and then download zip

It has all the following links, but it doesn't super explain it

You're going to need two models, you want the Full EMA I believe, though someone will tell me the differences I'm sure.

https://huggingface.co/CompVis/stable-diffusion-v-1-4-original

https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth

Put this file in the same folder as the files from the git link So it should be in the same directory as all the files from stable diffusion. As in webui.py webui.bat and your model file should all be in the same directory. Also, rename the 1.4 model you downloaded to model.ckpt. GFPGANv1.3.pth should remain the name it has.

You need to run and install python 3.10.6 https://www.python.org/downloads/windows/

You need to run and install git (64 git bit for windows setup) https://git-scm.com/download/win

and you'll need the Cuda Toolkit 11.3 (Windows, x86_64, version 10) https://developer.nvidia.com/cuda-11.3.0-download-archive?target_os=Windows&target_arch=x86_64

run webui.bat from Windows Explorer.

Just wanted to say thanks for this, I know nothing about python/git or anything, but I followed this and it's working perfectly for me!

Do you by any chance know of any simliar idiot's guide to textual inversion? I've looked at some tutorials but they're getting into editing code and such and it's going way over my head.

# ? Sep 24, 2022 16:14

mobby_6kl: Aug 9, 2009; by Fluffdaddy

Has anyone tried doing text inversion themselves yet? The Automatic1111 documentation refers to this guide: https://github.com/invoke-ai/InvokeAI/blob/main/docs/features/TEXTUAL_INVERSION.md

Which doesn't work with the Automatic build so I set it up in parallel. Had to downgrade torchmetrics to 0.6.0 and install some other random poo poo to get it to run. However it still fails after doing something for a minute or two:

code:

...
  File "C:\Users\xyz\AppData\Roaming\Python\Python39\site-packages\torch\cuda\__init__.py", line 211, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

What the gently caress? I literally just ran the SD text to image with no problems. Any ideas how is this possible?

Actually using the example embeddings works fine but I want to train it on my own dumb face.

# ? Sep 24, 2022 16:31

lunar detritus: May 6, 2009

mobby_6kl posted:

Has anyone tried doing text inversion themselves yet? The Automatic1111 documentation refers to this guide: https://github.com/invoke-ai/InvokeAI/blob/main/docs/features/TEXTUAL_INVERSION.md

Which doesn't work with the Automatic build so I set it up in parallel. Had to downgrade torchmetrics to 0.6.0 and install some other random poo poo to get it to run. However it still fails after doing something for a minute or two:
code:
...
  File "C:\Users\xyz\AppData\Roaming\Python\Python39\site-packages\torch\cuda\__init__.py", line 211, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled
What the gently caress? I literally just ran the SD text to image with no problems. Any ideas how is this possible?

Actually using the example embeddings works fine but I want to train it on my own dumb face.

Are you running it in the same environment as the automatic script? If you have been using the bat to run it, it's kinda hidden that it actually creates its own python environment with its own copy of torch and the like.

# ? Sep 24, 2022 16:34

EVIL Gibson: Mar 23, 2001; Internet of Things is just someone else's computer that people can't help attaching cameras and door locks to!; Switchblade Switcharoo

Comfy Fleece Sweater posted:

https://www.engadget.com/nvidia-ai-model-get3d-virtual-worlds-objects-130021127.html

3d is coming

https://www.youtube.com/watch?v=KfcsdhGKb1U

Nvidia releases tool.
Tool is, as per Nvidia quality, crashing every fifth time it's opened.
Find out you need to work with a very specific file format that records every fart you squeaked out when you brace to see it will actually save correctly.
Nvidia made "TressFX" and now no one uses it. Remembering this fact you cry when you have been tricked yet again!

# ? Sep 24, 2022 17:21

Attack on Princess: Dec 15, 2008; _{To yolo rolls! The cause and solution to all problems!}

When you reference an artist, does the AI imitate pictures from their library so to say, or does it just invoke their general art style?

Because if it uses artists like libraries, it's getting an increasingly large mountain of Greg Rutkowski material. A recursive loop of ever more detailed Greg art.

# ? Sep 24, 2022 17:49

Tunicate: May 15, 2012

Just a general vibe

# ? Sep 24, 2022 17:55

Comfy Fleece Sweater: Apr 2, 2013; You see, but you do not observe.

EVIL Gibson posted:

Nvidia releases tool.

Tool is, as per Nvidia quality, crashing every fifth time it's opened.

Find out you need to work with a very specific file format that records every fart you squeaked out when you brace to see it will actually save correctly.

Nvidia made "TressFX" and now no one uses it. Remembering this fact you cry when you have been tricked yet again!

I don�t disagree, it�s just the first AI text-to-3D I�ve seen in the wild. I wonder who else might be working on 3D AI.

I figured someone was probably hard at work on the idea, but didn�t expect it so fast (well, it�s still vaporware, but now I�ve no doubt it�s coming, if not from them, from someone else)

I was just getting into 3D modeling, now I feel like it�s just a matter of time before modelers get a kick in the balls like illustrators did :v:

# ? Sep 24, 2022 18:08

Fuck You And Diebold: Sep 15, 2004; by Athanatos

https://twitter.com/TomLikesRobots/status/1572625296481996800
https://twitter.com/paultrillo/status/1572301755710582784

# ? Sep 24, 2022 18:58

mobby_6kl: Aug 9, 2009; by Fluffdaddy

lunar detritus posted:

Are you running it in the same environment as the automatic script? If you have been using the bat to run it, it's kinda hidden that it actually creates its own python environment with its own copy of torch and the like.

Yes I was! I still don't understand why it wasn't working but whatever, I created a new environment and it was fine. Eventually I got to the point where it runs out of memory.

code:

RuntimeError: CUDA out of memory. Tried to allocate 256.00 MiB (GPU 0; 8.00 GiB total capacity; 6.16 GiB already allocated; 286.00 MiB free; 6.25 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Do you know how much is actually needed? Seems like 256MB might be doable if I could improve the fragmentation situation but I also don't want to spend the night loving with it only to discover it needs 20GB or something :v:

Alternatively, is there a collab I could just use for this?

# ? Sep 24, 2022 18:58

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 02:42

Fuck You And Diebold: Sep 15, 2004; by Athanatos

mobby_6kl posted:

Yes I was! I still don't understand why it wasn't working but whatever, I created a new environment and it was fine. Eventually I got to the point where it runs out of memory.
code:
RuntimeError: CUDA out of memory. Tried to allocate 256.00 MiB (GPU 0; 8.00 GiB total capacity; 6.16 GiB already allocated; 286.00 MiB free; 6.25 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Do you know how much is actually needed? Seems like 256MB might be doable if I could improve the fragmentation situation but I also don't want to spend the night loving with it only to discover it needs 20GB or something Alternatively, is there a collab I could just use for this?

I found that if I tried something that requested too much memory it would error out but never free up the memory it grabbed so I couldn't run anything else after that, had to quit out, restart it and try something a little less resource intensive.

# ? Sep 24, 2022 19:02

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

«‹›435 »