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

[The Mass Shooter Ron Soye]

Senate Intelligence Authorization Act advances unanimously out of committee (archive)

Provisions included:
* Protecting important voluntary investments in watermarking and content authenticity by AI firms by creating penalties for those that deliberately push removal of those voluntary protections.

I'll advocate for the removal of watermarks all day every day. Am I going to prison? Let's see what the bill actually says.

The relevant text in the legislation is in Section 511. PROTECTION OF TECHNOLOGICAL MEASURES DESIGNED TO VERIFY AUTHENTICITY OR PROVENANCE OF MACHINE-MANIPULATED MEDIA.

Spoiler

(b) Prohibitions.—

(1) PROHIBITION ON CONCEALING SUBVERSION.—No person shall knowingly and with the intent or substantial likelihood of deceiving a third party, enable, facilitate, or conceal the subversion of a technological measure designed to verify the authenticity, modifications, or conveyance of machine-manipulated media, or characteristics of the provenance of the machine-manipulated media, by generating information about the authenticity of a piece of content that is knowingly false.

(2) PROHIBITION ON FRAUDULENT DISTRIBUTION.—No person shall knowingly and for financial benefit, enable, facilitate, or conceal the subversion of a technological measure described in paragraph (1) by distributing machine-manipulated media with knowingly false information about the authenticity of a piece of machine-manipulated media.

(3) PROHIBITION ON PRODUCTS AND SERVICES FOR CIRCUMVENTION.—No person shall deliberately manufacture, import, or offer to the public a technology, product, service, device, component, or part thereof that—

(A) is primarily designed or produced and promoted for the purpose of circumventing, removing, or otherwise disabling a technological measure described in paragraph (1) with the intent or substantial likelihood of deceiving a third party about the authenticity of a piece of machine-manipulated media;

(B) has only limited commercially significant or expressive purpose or use other than to circumvent, remove, or otherwise disable a technological measure designed to verify the authenticity of machine-manipulated media and is promoted for such purposes; or

(C) is marketed by that person or another acting in concert with that person with that person’s knowledge for use in circumventing, removing, or otherwise disabling a technological measure described in paragraph (1) with an intent to deceive a third party about the authenticity of a piece of machine-manipulated media.

(c) Exemptions.—

(1) IN GENERAL.—Nothing in subsection (b) shall inhibit the ability of any individual to access, read, or review a technological measure described in paragraph (1) of such subsection or to access, read, or review the provenance, modification, or conveyance information contained therein.

It sounds like you could get in trouble for creating open source software that removes AI watermarking and provenance metadata. Removing some words: "No person shall knowingly and with the... substantial likelihood of deceiving a third party... facilitate...the subversion of a technological measure designed to verify the authenticity, modifications, or conveyance of machine-manipulated media, or characteristics of the provenance of the machine-manipulated media, by generating information about the authenticity of a piece of content that is knowingly false."

But there is that part at the end that says "by generating information about the authenticity of a piece of content that is knowingly false". If you simply remove it without substituting in any false information, maybe you're in the clear. In other words, you facilitate the removal but you don't make any claims about the authenticity or origin of the AI-generated media, so you aren't defrauding anybody.

Maybe content will be algorithmically suppressed by social media if it doesn't conform to "voluntary" authenticity standards, creating an incentive for users to either play ball or generate the false information, opening themselves up to potential civil or criminal penalties.

This is something to keep an eye on. Note that there is a big exemption subsection for "libraries, archives, and educational institutions", which would include entities like the Internet Archive.

 

[macrodegenerate]

Boy Oh Boy, I sure hope there is some kind of remedy or consequence for knowingly suing a person hosting legitimate footage under this act. Otherwise this just looks like a way for politicians and celebrities to silence opposition through the courts.

 

[Looney Troons]

Man, this is weird. I was just talking about this a few days ago.

A friend of mine is a photographer. Her largest clients are schools nearby her. Where the school pays her for her services, her bread and butter comes from parents buying photo packages, and especially digital RAW files.

On her website, she allows for potential customers to preview and select the pictures they’d like to purchase. Plastered over each and every low resolution picture are tons of egregious watermarks.

Edit: forgot the meat of this exposition. She has experienced a drastic drop in digital sales over the past two years and she attributed it to photo manipulation. To give you a taste of her pricing, a set of twelve digital photos in their RAW form costs 220 USD. You can buy the Adobe Creative Cloud suite for as low as ~40 dollars per month and read a brief tutorial to remove watermarks, and the Topaz Labs Gigapixel software for 100 USD to easily upscale the image… saving about 80 USD and giving you tools to play around with.

There are tons of utilities on Git et al that, I believe, use some sort of impainting technique to remove the watermarks it identifies. I use a combination inpainting and Photoshop to remove and unwanted text from images I generate… but I’m curious if commercial tools from Adobe or Topaz which can definitely be used to remove these watermarks will fall under the same scrutiny?

 

[Looney Troons]

Sorry for the double-post. I haven’t seen ToonCrafter mentioned.

“ToonCrafter can interpolate two cartoon images by leveraging the pre-trained image-to-video diffusion priors.”

You can take two images, give them a prompt, and it’ll spit out a seemingly seamless two second animation at 320x512 resolution. I haven’t given it a go on my rig yet, but it looks like it works surprisingly well.

 

[Baraadmirer]

Diving into my first foray with SDXL (particularly PonyXL) and I'm surprised that quite a few characters have been baked into the checkpoint.

 

[crowabunga]

Diving into my first foray with SDXL (particularly PonyXL) and I'm surprised that quite a few characters have been baked into the checkpoint.
View attachment 6065229

This reminds me, does anyone know anything about PonyXL drama? I recall hearing that the model's developer obfuscated a bunch of tags, and also won't share those obfuscated tags with people, but I expect I'm getting the technical details wrong.

 

[Looney Troons]

As far as I know, the “scoring” tagging goes like:

score_9 = high quality illustrations (akin to using masterpiece and best quality)

score_8_up = realistic texture and backgrounds

score_7_up = makes the image more anime-like

There’s so much drama in this community and they’re all just merging checkpoints off of existing ones without giving credit anyway.

 

[Baraadmirer]

As far as I know, the “scoring” tagging goes like:

score_9 = high quality illustrations (akin to using masterpiece and best quality)

score_8_up = realistic texture and backgrounds

score_7_up = makes the image more anime-like

There’s so much drama in this community and they’re all just merging checkpoints off of existing ones without giving credit anyway.

I wasn't aware that the score tags were that specific. This article ( Archive ) only says that those are required because of a training issue.

If you want some drama, check out this comment thread ( archive not available ) on AutismXL, where a user white-knights against the offensive name and is told off by some autists:

 

[AmpleApricots]

I've not looked more closely at imagen in a while but what happened to that alternative architecture by Stability, the big one? Bratwurst or something. When I tested it it was very promising. Did people actually do anything with it? There also has been a ton coming out of China right now. It feels a bit like all community progress halted. What happened? Did everyone just collectively decide to wait for SD3?

 

[Looney Troons]

I wasn't aware that the score tags were that specific. This article ( Archive ) only says that those are required because of a training issue.

If you want some drama, check out this comment thread ( archive not available ) on AutismXL, where a user white-knights against the offensive name and is told off by some autists:
View attachment 6066623

I forget the entire explanation that I was told, but you’re correct. There was some big mistake in the training data but the original model was put up because it worked really well anyway. I like all of the Pony models that I’ve tried if I’m generating artistic looking environments. I also forget what the specific ‘score down’ tags were supposed to do. I think 5 has something to do with appendages.

Someone is trying to create a no score version of Pony. I have the model downloaded (I’m collecting interesting ones and throwing them on the NAS), but I haven’t tried it yet. Description says it works without the scoring tags, but performs better with them... so, why not just throw them in?

CivitAI’s comment section is an absolute trove of stupidity. Tons of the “please don't use abelist language” shit, then when you check their profile out, they’ve uploaded nothing but bestiality porn.

 

[macrodegenerate]

As far as I know, the “scoring” tagging goes like:

score_9 = high quality illustrations (akin to using masterpiece and best quality)

score_8_up = realistic texture and backgrounds

score_7_up = makes the image more anime-like

There’s so much drama in this community and they’re all just merging checkpoints off of existing ones without giving credit anyway.

Majorly incorrect. Pony is trained off danbooru tags and metadata. The score tags are the scores from danbooru. Score_7_up is 7 score_9 is a score of 9 or higher etc.

 

[inception_state]

https://old.reddit.com/r/comfyui/comments/1dbls5n/psa_if_youve_used_the_comfyui_llmvision_node_from/

Infostealer malware found in a comfyui node called ComfyUI_LLMVISION - runs a keylogger .exe and then uses a Discord webhook to exfil the data. This is unfortunately quite easy to add to an extension, since you are basically downloading and running arbitrary code from the internet, sometimes in non-obvious ways. In this case the malware was built into a sub-dependency of the extension, custom wheels of particular libraries:

xxxx://github.com/AppleBotzz/Backup-Anthropic-Builds/raw/main/anthropic-0.26.1-py3-none-any.whl #Custom wheel cuz buggy

xxxx://github.com/AppleBotzz/Backup-OpenAI-Builds/raw/main/openai-1.30.2-py3-none-any.whl #Also Custom wheel cuz buggy

 

[macrodegenerate]

https://old.reddit.com/r/comfyui/comments/1dbls5n/psa_if_youve_used_the_comfyui_llmvision_node_from/

Infostealer malware found in a comfyui node called ComfyUI_LLMVISION - runs a keylogger .exe and then uses a Discord webhook to exfil the data. This is unfortunately quite easy to add to an extension, since you are basically downloading and running arbitrary code from the internet, sometimes in non-obvious ways. In this case the malware was built into a sub-dependency of the extension, custom wheels of particular libraries:

Imagine Gooning so hard on SD that you get your identity stolen and network owned.

 

[inception_state]

Imagine Gooning so hard on SD that you get your identity stolen and network owned.

To be fair, it's a pain in the ass to run a SD setup locally in a secure way. I guess you could deploy everything in a VM or a Docker container, but your average user struggles just to get it running in the first place. Then there are GPU passthrough issues, drivers, etc.

Edit: Here's a Windows guide for setting up a Docker container for Comfy: https://old.reddit.com/r/comfyui/comments/1dc80al/installing_comfyui_in_a_docker_container/

 

[Susanna]

To be fair, it's a pain in the ass to run a SD setup locally in a secure way. I guess you could deploy everything in a VM or a Docker container, but your average user struggles just to get it running in the first place. Then there are GPU passthrough issues, drivers, etc.

Edit: Here's a Windows guide for setting up a Docker container for Comfy: https://old.reddit.com/r/comfyui/comments/1dc80al/installing_comfyui_in_a_docker_container/

I guess docker is more difficult on Windows, but my experience from Linux is that docker is by far the easiest way to set up a local SD. You don't have to worry about any of the python breakage and package hunting, because everything is version controlled and standardised.

 

[macrodegenerate]

I guess docker is more difficult on Windows, but my experience from Linux is that docker is by far the easiest way to set up a local SD. You don't have to worry about any of the python breakage and package hunting, because everything is version controlled and standardised.

The main issue is that docker compose isn't supported. On linux after you've installed docker with compose and the nvidia container drivers, it's as easy as:

#!/bin/bash
#divine-intellect.sh
if ! test -f /opt/ComfyUI/requirements.txt ; then
    apt-get update
    apt-get install -y curl git python3 pip
    cd /opt
    git clone https://github.com/comfyanonymous/ComfyUI.git
    cd /opt/ComfyUI
    pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121
    pip install -r requirements.txt
fi
python3 /opt/ComfyUI/main.py --listen

#docker-compose.yml
services:
  comfy-ui:
    image: nvidia/cuda:12.1.1-runtime-ubuntu22.04
    volumes:
      - ./ComfyUI:/opt/ComfyUI
      - ./divine-intellect.sh:/opt/divine-intellect.sh
    command: "chmod +x /opt/divine-intellect.sh; /opt/divine-intellect.sh"
    ports:
      - '8188:8188'
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

docker compose up

Windows not so much. Although supposedly you could get that functionality by installing the compose plugin. Also they have a portable windows download for comfy.

To be fair, it's a pain in the ass to run a SD setup locally in a secure way. I guess you could deploy everything in a VM or a Docker container, but your average user struggles just to get it running in the first place. Then there are GPU passthrough issues, drivers, etc.

Edit: Here's a Windows guide for setting up a Docker container for Comfy: https://old.reddit.com/r/comfyui/comments/1dc80al/installing_comfyui_in_a_docker_container/

There's no point in trying to run a secure VM on windows using WSL. I think the VM will always have access to your file system. The best you can do is run it through virtualbox, but passthrough isn't available like on KVM/QEMU using PCIe passthrough.

To be truly secure your best bet would be to run it on another machine on your network and ban all outgoing traffic from that machine besides port 22(SSH) and the application port.

Edit: Niggerlicious

 

[Overly Serious]

The main issue is that docker compose isn't supported. On linux after you've installed docker with compose and the nvidia container drivers, it's as easy as:

Say what now? Windows has Docker Compose. Of course it does - I use it all the time.

There's no point in trying to run a secure VM on windows using WSL. I think the VM will always have access to your file system. The best you can do is run it through virtualbox, but passthrough isn't available like on KVM/QEMU using PCIe passthrough.

It has whatever permissions it was given. WSL is Windows Subsystem for Linux and version 1 wasn't a VM at all - it was a compatibility layer that let you execute Linux API calls to the Windows kernel. WSL2 does run in a VM on Hyper-V and has an actual Linux kernel. However, it is still by default running using permissions of its creating host. So if you as an authenticated user want to create an Ubuntu instance under that user profile, it can access the host as that user but not as other user accounts on the same machine (if they exist). That's the intent. If you want a separate VM you just roll one up with Hyper-V. Docker containers using WSL2 backend have the permissions local to the container and those provided by Docker environment. I.e. if you bind mounted a directory into a container then yes, it could write to that directory, but it's not going to jailbreak itself and do anything it wants on the host unless the host gives it that ability. And you could, if you wish, create separate account and ACLs for Docker to execute under. But you could also just create a VM. Which doesn't require Virtualbox, you'd just use Hyper-V natively.

Personally I run Stable Diffusion on an Ubuntu OS via dual boot because I have an AMD card and ROCm for Windows is very, very far behind. You can't (remotely easily) do pass-thru to the card via WSL2.

 

[AmpleApricots]

Imagine Gooning so hard on SD that you get your identity stolen and network owned.

This honestly can happen in tons of software. Millions of dependencies, very little oversight what's even in them, all it takes is one github to be compromised or one developer with a hidden agenda. These are called supply chain attacks, not exactly what happened here but comfyui also has a rats nest of dependencies. I would always sandbox applications like Comfy. The open source community is far too trusting to run abritary code. The amount of times some anonymous, random literally who just links his github on reddit and says "hey people run this" and people actually just do is insane.

 

[Joshua Connor Tomar]

Some images I generated with AI.
View attachment 6056075

How'd you make the second one? Love it.

 

[Alex Hogendorp]

How'd you make the second one? Love it.

Keep in mind I remixed it from a similar portrait on Seaart.

(masterpiece, best quality:1.2),intricate details, solo, armor,jakubrozalskistyle warhammer, 40k, imperial, guardsman