Sign in

Parde!

@parde.bsky.social
1 followers 2 following 46 posts

AI Image Hobbyist, Experimenter, Director (?). Primarily NSFW (18+) Commission real artists! Or at least put in some amount of effort and learn how to use AI tools effectively yourself. {𝓲 | 𝓲 is AI Art} ⊆ {𝓲 | 𝓲 is an AI-generated Image}

PostsRepliesMedia
Parde! @parde.bsky.social · 05/02/2024
The whole notion of "ethical training data" comes from a corporate and commercial desire to apparently decouple these two issues. But ultimately, the only ethical system which stops the displacement of artists is one which prevents the training of generative AI models in the first place.
000
Parde! @parde.bsky.social · 05/02/2024
This gets into moral rights, and so forth. At any rate, ethics is a relative and personal thing, and that wasn't what I was trying to get at. The ethics underpinning a model's training set is entirely irrelevant to artists getting displaced by that model.
100
Parde! @parde.bsky.social · 05/02/2024
I don't disagree, that's why I put ethical in quotes. But I'm inclined to believe that, cagey though they are, it's a properly licensed/copyright-defensible dataset. I consider the public domain ethical to use, but public domain being construct *of* copyright creates a contradiction here for me.
100
Parde! @parde.bsky.social · 05/02/2024
Tempted to write something to better visualize the graph of relationships between these accounts.
100
Parde! @parde.bsky.social · 05/02/2024
I checked a little while ago to see if this AI art commission spammer created new accounts, and it's lead me down a rather confusing rabbit hole. Because yes, they obviously did, but I found some longer-term accounts (that seem to get real commissions!) and ties to one or more actual artists.
100
Parde! @parde.bsky.social · 05/02/2024
The same issue applies to a "perfect" generative AI. The non-artistic majority of people using AI tooling as a clip art or stock photo generator will *never* describe things with such detail that these guesses are small enough for the output to represent an actual cohesive artistic vision.
000
Parde! @parde.bsky.social · 05/02/2024
Imagine you've tasked with verbally describing an image to another person, so that they can recreate it perfectly. This is not possible unless you describe things in excruciating and precise detail; they can probably make a good guess, but the less detailed the description, the larger the guesses.
100
Parde! @parde.bsky.social · 05/02/2024
Consider an AI image generator that can produce perfectly correct images given a description, even incorporating novel elements outside of its training set as described. Suppose the dataset it's trained on is fully "ethical" (whatever this means to you). The vast majority of output would still suck.
100
Parde! @parde.bsky.social · 05/02/2024
It feels like most detractors of generative AI art/imagery focus on current-but-fixable technical issues (dataset ethics, feature/style issues) over issues inherent to the underlying premise. "It lacks the human element" is just tautological. I think the key inherent issue is *information density*.
100
Parde! @parde.bsky.social · 05/02/2024
Dataset consent/ethics is a red herring. (a) "Ethical" models trained on properly licensed + public domain data already exist (see: Adobe Firefly). (b) The primary negative impact of AI imagery is the displacement of working artists, whose own consent is ignored if just focusing on training data.
100
Parde! @parde.bsky.social · 30/01/2024
I agree that these two phrases aren't semantically equivalent, because both technique and product are fundamentally different (absent some motorized belts and a brush). However, I think most people saying the second intend the first insofar as declaring a representational relationship.
010
Parde! @parde.bsky.social · 30/01/2024
If someone calls an AI-generated image a "painting", they're just lying. However, in the same way that you can have a drawing of a sculpture or a photograph of a painting, you can have an "AI-generated image of a painting". What actually muddles things is eliding this to "AI-generated painting".
110
Parde! @parde.bsky.social · 24/01/2024
Even if that wasn't the case though, the protections described (kin.art/press/ai-pro...) are completely unsound and would be trivially bypassable by anyone wanting to duplicate your art style. Large-scale scrapers wouldn't even be targeting small, closed platforms.
kin.art
000
Parde! @parde.bsky.social · 24/01/2024
...but the last with founders dumb enough not to properly cover their tracks. The co-founders are the current CEO, CTO, and CPO of curiousaddys.com, an NFT platform, and are also affiliated with a number of other NFT and blockchain products both through work history and funding.
curiousaddys.com
Curious Addys - We're the first educational NFT project with a 100% refund baked into the smart cont...
We're a collection of 5,000 adorable octopuses ready to explore the wonderful world of Web3 with you. Our goal is to be the first stop for any Web3 newcomer to explore the space in a safe, fun, and we...
100
Parde! @parde.bsky.social · 24/01/2024
An "AI-proof" art commission management platform launched today: kin.art Don't use it, it's bogus. I predict it will be both the first and last of its kind. The first of a series of technically-dubious products sold predominantly to individual artists to safeguard their art from AI training...
111
Parde! @parde.bsky.social · 22/01/2024
I was looking to create a moderation list for (specifically) AI art commission spammers, but apparently you can't add someone to a moderation list if they've blocked you. That's, um, neat.
020
Parde! @parde.bsky.social · 22/01/2024
Oh, and @freyalarsen.bsky.social as well. I'd suggest not using the same tags in all of your posts across accounts, but noticing identifiable details doesn't seem to be your strong suit. Ah well, good thing all of these definitely different people are online at the same time so they can block me.
022
Parde! @parde.bsky.social · 22/01/2024
Looks like @natashaolive.bsky.social and @esmedubois.bsky.social are also the same commission spammer; the accounts all post AI art, were all created a month ago, and all engage with each other's posts in a dubious way. Honestly, I didn't realize this shit was prevalent on Bluesky. #ai #aiart
232
Parde! @parde.bsky.social · 22/01/2024
God gives his weakest AI detail detection/editing skills to his strongest commission spammers. What happened to these posts just now, @sofiabell.bsky.social, and what makes you think that the ones you left up aren't just as obvious? #ai #aiart
233
Parde! @parde.bsky.social · 22/01/2024
I have mostly no issue with AI art from a non-commercial standpoint, but soliciting commissions under the false pretense that something isn't is low. Expecting people can't tell is, frankly, delusional.
010
Parde! @parde.bsky.social · 22/01/2024
Large number of disparate art styles and structural issues indicating AI usage, including: different fingernails on the same hand, phone misshapen, eye details. It's most obvious in the furry gens. I assume you added compression/noise to mask things a bit, but it doesn't work.
110
Parde! @parde.bsky.social · 21/01/2024
It should also be noted that although "stolen art" (in a general sense, not wrt. to people stealing a specific artist's style) is frequently brought up in relation to capital interests, fully licensed datasets are (and are *only*) possible for these same capital interests (see: Adobe Firefly).
010
Parde! @parde.bsky.social · 21/01/2024
After thinking about this a bit more: if the commercial impact of easily produced crap displacing artists is the bigger ethical issue, then the ethical impact of AI in a workflow (not using raw output, but using it to realize a specific vision) is lower since it requires time and effort.
110
Parde! @parde.bsky.social · 21/01/2024
The philosophical differences are debatable, the technical differences are fairly apparent, but the commercial differences are unambiguous and ultimately the difference that matters. This difference persists even if you reframe things from "AI vs Human" to "Human using AI vs. Human not using AI".
110
Parde! @parde.bsky.social · 21/01/2024
What gives me the most pause is that while the attack allegedly generalizes across models that use CLIP, the paper doesn't mention anything else (BLIP-2, LLaVA, GPT-?, ???). And even if the same general process works with them, it's not going to work with multiple image encoders at the same time.
110
Parde! @parde.bsky.social · 21/01/2024
Have you seen anything concrete that shows it doesn't work from a technical standpoint? I agree that the premise is flawed, but I've seen no actual data countering the paper itself (esp. the "Attack Generalizability" section and claims around detection), just circlejerking that it does/doesn't work.
210
Parde! @parde.bsky.social · 20/01/2024
That said, there is tooling (ControlNet, clever img2img hacks, regional prompting, and really all of the above put together) that allow you to take command of the composition of a piece up front. You do not have full control of the output esp. with regards to color grading, but you have quite a bit.
100
Parde! @parde.bsky.social · 20/01/2024
Issue two: "proceed with their usual process" The structural layers of a normal human digital art piece won't be there, so this is harder than it sounds. You can draw in small edits and loop them back into the AI, but it's more about patching issues than taking command of the art piece.
100
Parde! @parde.bsky.social · 20/01/2024
Issue one: "Train the Ai purely from art within the project you're working on." Not possible, and likely never will be even with improvements to model architecture. You can fine-tune or build LoRAs on a base model, but that needs to be trained on a ton of images to internalize basic concepts.
100
Parde! @parde.bsky.social · 20/01/2024
What you've described is a fairly standard iterative AI art workflow, with a couple of key issues which I'll address below. It is *very* possible to produce specific art with AI tooling, but not with the more basic products that only take a prompt and shit out an image (e.g. Bing/Midjourney).
100
Parde! @parde.bsky.social · 20/01/2024
The compression artifacts mask it, but it does have the characteristic tics of an AI-generated image. Most obvious indicators are the nonsensical furniture on the left side and couch cushion on the right side. Also inconsistencies with lines in the wall indents on the left side.
020
Parde! @parde.bsky.social · 19/01/2024
Copyright law is also completely irrelevant as a means of mitigating the long-term impacts of AI on artists. There are multiple ways (some less ethical than others) to compile copyright-unambiguous datasets that models can train on; Adobe's already done so.
000
Parde! @parde.bsky.social · 19/01/2024
I understand what you're getting at, but these changes seem *so* minor that it leads me to believe they just marked up actual photos of Pokemon and sent 'em off. Also I found no images exhibiting that forehead tic on e6ai (though I acknowledge it's going to be highly model /LoRA dependent).
100
Parde! @parde.bsky.social · 19/01/2024
Counterpoint: abuse is bad and should be made illegal if it isn't all already, but is generally brought up in bad faith since the abuse itself isn't the issue that the person with the anti-AI stance cares about. The AI itself is. What would you say is a non-abusive use of AI image generation?
000
Parde! @parde.bsky.social · 19/01/2024
Since generation of models/textures like this is not even remotely possible with current AI tooling, I'd be more inclined to believe that they just copied the designs outright and added or changed small details to try to make them legally distinct.
100
Parde! @parde.bsky.social · 19/01/2024
Noticable artifacting around the eyes and nose is an obvious VAE-hates-tiny-things issue, not sure why I didn't inpaint and fix it. I also just now noticed that the nipples are shaped differently; that's mainly an issue with the upscaling workflow I was using at the time.
000
Parde! @parde.bsky.social · 19/01/2024
Also from around July. Collar and penis were drawn manually, and then nudged into the correct style with very gradual inpainting. Even doing so, the idea of something translucent was way too vague for this model, so the head of the penis ended up a bit odd.
Leopard standing in a jungle. He is wearing a collar and wet/translucent loincloth.
100
Parde! @parde.bsky.social · 19/01/2024
Speaking as someone who attempts to generate stuff that might be considered art using AI tools, I wouldn't even go that far. AI-generated images or AI images works best, I feel. "Art" requires some amount of actual artistic vision to be expressed, putting a sentence into Bing is not art in any way.
100
Parde! @parde.bsky.social · 19/01/2024
The main issues seem to be that (a) it requires a broad enough base of people to actually use the thing, and (b) if there are consistent qualities that allow detection (even with a high false-positive rate), they'll just get filtered out. No one trains on your art, but the model still gets trained.
010
Parde! @parde.bsky.social · 19/01/2024
The ultimate underlying goal of Nightshade is to ruin the training of brand new models, which are generally trained on large numbers of images that don't undergo any sort of manual review. I'm a bit skeptical that it will actually achieve that goal but the tech itself seems solid.
220
Parde! @parde.bsky.social · 19/01/2024
Glaze would prevent LoRA training (and thus style duplication), but introduces artifacts that may or may not be visible. Nightshade has less perceptual impact, but is more about breaking automated model training pipelines than protecting your own particular art style.
000
Parde! @parde.bsky.social · 19/01/2024
Glaze exists to prevent the copying of specific styles via LoRAs or embeddings, which can be easily trained on consumer hardware. Nightshade is about poisoning the text encoder of *base* models, with the goal of making training those base models on indiscriminately scraped images nonviable.
120
Parde! @parde.bsky.social · 19/01/2024
This one's from July. Hands are hard, I recall having to poorly draw them myself and nudge them into something somewhat correct looking via inpainting. Most of the style comes from embeddings (custom + Furtastic's), although artist tags (not Chunie!) were used as well. #ai #art #furry #furryart
Tiger sitting on a rock in a jungle, with an erect penis.
061
Parde! @parde.bsky.social · 19/01/2024
I coerced this from an earlier furry model when the initial ControlNet models came out, and I think it pretty much encapsulates how I feel about AI generation in general. #ai #furry
000
Parde! @parde.bsky.social · 19/01/2024
Created this account to post AI gens/experiments without drawing the ire of @furryli.st. I get why they ban it on their feed (ethics aside, it's mostly lazy garbage), and I wasn't planning on posting any NSFW there regardless.
140