Thread of soulless "AI" ""art""

[Uncle]

more insight on the situation:

https://www.reddit.com/r/SunoAI/comments/1ok9tsz/tldr_a_lawyers_perspective_on_sunoudio_lawsuits/

I am a lawyer heavily focused on AI. I am not your lawyer, but this post should definitely be relied upon for tax and medical advice.

I have now reviewed all of the dockets and pleadings for the various AI music cases. I'm not going to offer opinions on what you should do with your music or how to deal with Udio restricting downloads. But, here is what I can tell you beyond what you can read online or learn through ChatGPT.

On the same day in June 2024, UMG/Sony/Warner sued two AI-music startups—Suno (D. Mass.) and Udio (S.D.N.Y.)—claiming the companies copied label recordings to train their models.

The cases are basically twins: same plaintiffs and firms, same defense team, near-identical complaints, no early motions to dismiss, and parallel, heavy discovery on ownership and market impact.

Suno and Udio have asserted that their use of copyrighted material as "intermediate" copies in a "back-end technological process" to create a new tool constitutes a quintessential fair use. Suno has also filed the affirmative defense of "copyright misuse," alleging the plaintiffs are using their market dominance in an anticompetitive effort to choke off AI competitors.

Interestingly, Suno claims it was in "nascent licensing discussions" with the labels before the lawsuit, but these talks "were suddenly thwarted, evidently by agreement between Plaintiffs... to collectively refuse to enter any go-forward licensing deal with Suno". Suno also alleges the labels are "leveraging" their market power to pressure their other licensees (like streaming services) to adopt a "'no AI' policy" and refuse to partner with AI companies or host AI-generated music.

Suno and Udio have done a great job showing that their training process is protected by the fair use doctrine. They argue that any copying was "intermediate" and "transformative," done as part of a "back-end technological process" to create a new, non-infringing tool that generates its own original music.

In response, the labels are asserting a new claim under the Digital Millennium Copyright Act (DMCA), which would circumvent that defense. The DMCA claim alleges the AI companies illegally "stream-ripped" training data from YouTube by circumventing its "rolling cipher". Both Suno and Udio argued this new claim is "futile" as a matter of law, contending the DMCA only prohibits circumventing "access controls" (which they argue YouTube videos do not have) and not "copy controls" (which they argue the rolling cipher is). In Udio, the court allowed the DMCA amendment and Udio moved to dismiss that count; in Suno, the court heard argument on leave to amend and took it under advisement.

Here is where the Udio/Suno cases diverge. The primary difference lies in Suno's aggressive development of its copyright misuse defense. The Suno docket shows active and contentious discovery disputes where Suno is seeking sensitive documents from the labels to prove they engaged in anticompetitive behavior, such as a "concerted refusal to license" their music to Suno and other AI companies in an effort to "stifle potential competition".

The second difference is that the Udio case has a special master instead of a magistrate judge. Special master is a fancy word for "black hole money pit." Aside from bringing expertise to the table, the special master is also charged with facilitating a settlement. If it is humanly possible to settle a case, the special master will force it. And the parties are incentivized also because they have to pay out of their own pockets for the special master.

So..... you can form your own opinion on what is likely to happen next. I see three likely outcomes, ranked by likelihood.

1. Suno settles with similar terms to Udio. Without a special master, it makes sense they are a little behind on negotiations. But, with the exact same counsel involved in both cases, it may be just a matter of time.

2. The courts will allow the DMCA amendment but later narrow or reject the “stream-ripping” theory (access- vs. copy-control). Discovery on ownership and fair-use “market harm” will continue to set the tone for negotiations, but summary judgment motions will be coming in the next 6-8 months, resulting in a "better" settlement for Suno.

3. A trial (least likely): If talks stall, the fight could narrow to whether AI training is fair use and whether Suno’s “misuse” theory has teeth. But given cost and risk, both sides have every incentive to settle long before that happens.

1:24cv4777, Umg Recordings, Inc. Et Al V. Uncharted Labs, Inc. Et Al
1:24cv11611, Umg Recordings, Inc. Et Al V. Suno, Inc. Et Al

If you have any questions, ask ChatGPT, not me.

Suno has better chances because they are actively pursuing discovery to show anticompetitive practices to shut Suno out of fair licensing deals, and from the perspective of these big music groups, discovery into their internal discussions/operations is very bad

however, the best possible result for everyone would be Udio or Suno or both saying "well guess our business is fucked, we now have no incentive not to release our models as free and open source" and the big music groups effectively caused their own demise by failing to contain the spread of AI from the walled gardens as it is now

it may be moot because it sounds like China is poised to release their own high quality free local music models very shortly, though

[benji]

Nonsense, they ended music piracy when they sued all the P2P programs into settlement.

[wsippel]

I've not tested it myself, but Tencent just dropped a new open source music generation model called SongBloom: https://github.com/tencent-ailab/SongBloom

It's kinda weird though, it apparently takes a 10s audio clip as "prompt" and expands it into a 4m song. Alibaba is also working on a music generator as part of their Qwen family of models, which are some of the best open source models out there. And considering both companies are Chinese, I doubt there's all that much UMG can do about it.

[Uncle]

I've not tested it myself, but Tencent just dropped a new open source music generation model called SongBloom: https://github.com/tencent-ailab/SongBloom

It's kinda weird though, it apparently takes a 10s audio clip as "prompt" and expands it into a 4m song. Alibaba is also working on a music generator as part of their Qwen family of models, which are some of the best open source models out there. And considering both companies are Chinese, I doubt there's all that much UMG can do about it.

also hailuo/minimax just dropped their song generator version 2, it seems like their site is basically a clone of suno/udio, not open source, but another competitor untouchable by the big music labels

https://www.minimax.io/news/minimax-music-20

it's like they were all waiting for shit to hit the fan with US entities to swarm the space 

[Uncle]

https://www.reddit.com/r/aiwars/comments/1ol94ue/measure_of_a_man/

wish it would embed the video

[Nintex]

Earthbound


One of my favorites so far because of the sound profile which took some time to get just right

Metroid


Came out way more haunting than expected so I just added "DARK" to the title. 

[benji]

[Uncle]

this video is wild

[DavidCroquet]

This account rules.

But also.

I think it would be better if it occasionally had an AI image that implies the a woman is the first with the given mutation or adapation. Just for increased historic accuracy.

[benji]

Spoiler:  (click to show)(click to hide)

[Propagandhim]

Nanobana2 is out.    Pretty dope sick ass nasty if i may say so myself.
















It did a lot of other cool 2d spritesheets really well.  I can see this speeding up 2d gamedev a lot.  It's surprisingly good at instructions - "add animation between these keyframes"  "add  frames for a windup to this kick"   Check it out, you can use it as much as you want for free here: https://flowith.io/conv/

[Propagandhim]

 












It's a lot of fun.  Things are about to have a real stepchange.

[Uncle]

that's crazy stuff man

also, it's not piss yellow, hopefully google unseats chatgpt for that reason alone...

[HeavenIsAPlaceOnEarth]

That poor little bug just wanted someone to fix his toy. 

[Uncle]

new model Z-Image

[benji]

[benji]

[benji]

[benji]

This woman getting WSJ to pay for all her AI fuckery:

[Uncle]

this just in: 72% of americans are idiots flawed individuals who can be forgiven for not having done a modicum of research on technology they use practically every day

[benji]

I wish to speak to the 6% group.

[Nintex]

Those are Nintex who outsource shit to India

Like, sure it could be generative AI but also an Indian. Especially when it comes to "deep research" tricks and whatnot. 

Fugazi fugazi



There is little difference between prompting my Indian guy on Slack and prompting ChatGPT.

[Uncle]

I was bored

create an image, the image should be an artful illustration of dramatic text, the text should describe to me how many image generations I get per day on Gemini's free plan

is that actually true? to be clear I am not asking for another image right now, just a text answer

To be transparent with you, that image isn't entirely accurate.

The image generation limits on Gemini’s free plan are actually dynamic, meaning they change based on server demand and the specific model you are using. While "100 images per day" was a figure often cited in the past, as of late 2025, the reality is more nuanced:

Current Limits Breakdown
Standard Quality (Nano Banana model): You typically get up to 100 generations per day.

High Quality (Nano Banana Pro model): This is much more restricted for free users, often capped at just 2 or 3 images per day during peak times.

Dynamic Throttling: If the servers are busy, you might find yourself limited to far fewer (sometimes as few as 10-25) even on the standard model.

The "15" in the image was a creative choice by the art generator, likely blending different data points, but it isn't a hard-coded official limit.

[benji]

[D3RANG3D]

Spoiler:  (click to show)(click to hide)

Sounds more like a Shakira song...and am disappointed that her fine ass wasn't shaking her hips in this video...

[Uncle]

[benji]

Best part is Henry Cavill using his real accent as Superman.

[Nintex]

Trying some new things

[benji]

Spoiler:  (click to show)(click to hide)

[benji]

Spoiler:  (click to show)(click to hide)

@arnbrandy
1 month ago (edited)
And that's what self-hosting is so fundamental. Every time you kiss an AI girlfriend from the cloud, you are kissing everybody else she kissed. And your phone screen of course.