Official Introducing Unity Muse and Unity Sentis, AI-powered creativity

Indeed. That a model contains foreign IP like Mario, has little to do with whether the images themselves are copyrighted (which is what Steam is "worried" about). Trademark (IP) and copyright are not the same.
https://copyrightalliance.org/faqs/difference-copyright-patent-trademark/

 

Given that Unity filled in the gaps with Stable Diffusion, I'm not surprised this happened. Also wow, these look like F***ing garbage.

 

They're different aspects of intellectual property but you can have both of them on the same thing.

 

That was a Wikipedia article as a prompt - that's not how you normally prompt an image generator. You gotta at least describe you want a humanoid or something.
In this case it is GOOD that they came out crap because it shows they have not used something heavily building on Stable Diffusion.

Here in comparison, 3 of 10 random results from Stable Difusion for the same wikipedia prompt (the other 7 were mostly nonsense):

Of course, but I'd say it's plausible that Unity has some images of Mario with a license to use them (the images for training, not the IP in something published).

 

No, it's really not. You're making excuses for Unity rolling out a bad product again. These aren't the only results that look bad, even the results they got from their own demonstration look bad.

 

AI bubble is starting to pop. The profitability is just not there in consumer markets. We already had reports of AI companies taking massive losses per user for growth purposes in hopes compute will get cheaper in the future and they'll be able to optimize the models. Unity has jumped on another defunct hype train as usual, same S*** as Weta.

Unity's audience for generative AI is already incredibly limited. It'll be interesting to see if they can even sustain this. Albeit, they seem to be using cheap models with lower compute requirements as demonstrated by the S*** results. Offering a more expensive service with worse results is not a winning formula.

 

Lol, okay bud..

 

I generally agree with you, except for this one.

WTF? VFX is "a S*** hype-train" now? If you said VR/XR/AR/Metaverse I would understand and agree, but thinking that a multi-billion business field: motion-picture and real-time VFX is a S***ty hype-train. ROFL. You don't know what you're talking about.

 

Unity got onto the Hollywood hype train a year or two after Unreal did, paid US$1.625B and has 0 to show for it. I'm not making any claims about other industries.

Also, Unity acquired only the tool making part of Weta - Weta Digital. They're not actually making any movies. Wētā FX, the actual VFX studio, is still headed by Peter Jackson and has no direct relation to Unity.

Also #2, the recent Weta Tools rebrand is 99.9% Ziva and Speedtree. They're trying to make it look like Weta has resulted into something, when it in fact it hasn't at all beyond a couple of Nuke and Maya plugins that are not available yet. That's exactly 0 synergy with the core engine product and few niche plugins are not worth US$1.625B.

 

Oh. Sorry I brought anything up. Forget me.

 

Who else has this when trying to access muse.unity.com ?

 

Try https://muse.unity.com/en

But yes, seems /fr does not work

 

Unity Muse may get potentially and coincidentally lucky, considering recent event of open AI, and now unclear future for ChatGPT.

 

Well, Muse is confirmed to be using ChatGPT under the hood so they are very reliant on it.

 

Ah, I have missed that one. Although I do remember someone has mentioned here this already.
Well then. That is even worse at the current state. We will see, which way wind going to blow.

 

But it's not that they use the OpenAI backend, right?
They surely have bought a model which was then trained on Unity resources. Selling custom trained models has been OpenAI's business for years before their breakthrough of ChatGPT after all.

However, CEO swaps or not, it would be everything but economic to stop their product as a whole. Why should that happen? Worst case is that further improvements become slower (and the competition catches up).

 

Cant be GPT4 since it consistently provides worse results. Likely 3.5 or one of the other cheaper models.
EDIT: Muse itself says it's GPT3 based:

I wouldn't be so sure, hardware for AI training is prohibitively expensive, especially for Unity who can't afford to bleed even more money. Same for the whole infra. My guess is they pay for a fine tuned GPT3.5 API access with a custom enterprise level deal.
...
Youtube recommended me this video, which explains why Muse's value proposition doesn't make any sense. The credits system on top of monthly sub means you can only do x120 Muse chat messages per month for 30$/month or you could skip Unity's subpar deal and get 9600 ChatGPT4 messages and unlimited ChatGPT3.5 messages for $20/month from OpenAI.

Furthermore, the credits system is centralized for all Unity AI services so if you gen some textures or sprites, you can send even less chat messages and vice versa. And all texture and sprite results are worse than what competitors offer for cheaper.

Spoiler: bad value proposition

 

@LeonhardP This looks awesome!

Will it be available to everyone or only to the paying members? And when?

 

https://store.unity.com/configure-plan/unity-muse-trial#:~:text=15-day free for first,access all of Muse's capabilities

And because the store page is intentionally deceptive in how it presents the pricing, and at a quick glance you might think it's free (or "$0 per 15 days" which is a flat out lie), it's actually 30$ per month after a 15 day free trial.

 

So I thought. Thank you for clearing this point though.
I never go to the plans page, I guess I would've seen this if I had.

 

Thank you man, I had to change my browser favorite language.

Now I have the following problem :

Am I cursed or what ?

 

No, Unity just makes buggy products with poor implementation and scalability. It's like their whole thing.

 

It looks great, I really like it.
Although I don't use any AI, because I doubt whether the copyright of it (something synthesized through AI) is legal.

 

Sigh. Just got an email encouraging me to upgrade to Pro and they will give me 1 free month of this "still in beta" service. Where is the value in that? I thought that they originally announced that part of pushing Plus subscribers up to Pro would include Sentis and Muse, but now it seems it's just the crumbs?

 

There is no added value, just a price hike after the current Plus pricing runs out in a year after which it's you'll have to pay x4 more for Pro than you did for Plus. The free month of AI is just to upsell you a new service, where you pay $30/month to get 3000 credits, which will net you 120 Unity Muse chat messages per month. Fewer chat messages if you generate some textures or sprites and vice versa, since credits are shared. A deal of the century, if you ask me.

 

Anyone is using Sentis?
On Sentis forum there is one AI student using it to do remarkable things that it wasn't designed for.

 

Either they went back on their word, I believe this was told in the original terms, so after they had to "walk back", they probably felt less inclined to add "additional benefits" to Pro.

Or they will do it, but only after the new terms go in effect with Unity 6?

In the end it doesn't matter, it seems dead in the water. Either one of the many lawsuits bursts the AI bubble, or Unity AI tools become eclipsed by the competition (what they showed seems super underwhelming).

In any case, I don't see how this is worth any amount of dollars per month.

 

Link?

 

why not make a dots-based behaivor tree rather than another gameobject behavior tree compete with asset publishers.

 

two questions
1. Any tech details ? for example, since we know DOD often do not use inheritance, how do we doing Custom Node thing? i saw something similar is phil's virtualobjects .
2. Any timeframe? i would very much like to use it now

 

Hi. I prefer dots utility AI since it's much better tooling and work much naturally with dots. Even better implement multiple different AI solutions that can mix and match. Official can start from dots utility AI and dots behavior tree.

 

Dude. It seems you need to do a deep dive into current AI technology. You expect it to read your mind and make perfect models of exactly what you want from... a wikipedia article as input??? You clearly have little understanding in the matter.
In fact, it seems the majority of posters here have very minimal understanding of AI as a whole, current models, their usage etc., yet feel they need to chime in with some expert opinion. lol Calm down kids. Watching a youtube video about AI doesn't make you knowledgable on the subject.

 

I'm not a dude and I promise you I have done far more research into AI than you have because it's part of my job. My job is to do tool evaluation and even Unity's own demonstrations show that their output is absolutely garbage and you'd know that I mentioned that if you read the very thing you're quoting.

 

Actually, I would argue the opposite: With Stable Diffusion, when using single generic names like "Ashley" or "Henry", or even fake/ambiguous names like "Doctor Butt Face" or "The Stomach Muncher" (both things I have actually used), it is still able to generate high quality (and humanoid) results. With the Mario example, the fact that the results from Muse look bad is not an indication that it didn't know what Mario was, but rather, that it is bad at generating him altogether.

In fact, if you want to test this, instead of using Mario's name directly, just use a generic description of him (i.e. a prompt that contains no intellectual property). For example, with LoraTheExplorer (Stable Diffusion XL using LoRAs), using the prompt "short man, red newsboy cap, mustache, red shirt, blue overalls, gold buttons, brown shoes" gives this:

Spoiler

p: short man, red newsboy cap, mustache, red shirt, blue overalls, gold buttons, brown shoes
n: glitch, deformed, mutated, ugly, disfigured
w: 1
l: 1-3: Watercolor Style; 4: 3D Render Style

As you can see, it is possible to generate high quality images of subjects that look like Mario without actually generating Mario himself. So, with Muse, if the reason the results are bad is because the model was not trained on Mario, then it should still deliver high quality results when given a generic description of him that is void of any IP.

 

@ShaneeNishry?

 

But something like your prompt is exactly not what was given - instead the content of Wikipedia was given which is anything but a proper visual description (the way AI generators need it).

 

Looks very promising but I do n ot understand why they have this subscription without the ability for the dev to have a real test period or a free version so we can start using it. I guess if it is as good as advertised we devs will be hooked if we start serious development with it. I am not going to use it due to this. At least not at this stage.

 

But if it's not good and they don't require you to commit to at least one paid month, they get nothing.

Now let's ponder why this is the actual route they took and what it says about the confidence they have in their product.

 

I know, that's the point. Try using my prompt in Muse and see if the results are better. If Muse is a good AI image generator, it should at least be able to generate an image of generic things; after all, my prompt is just clothing items with colors, nothing specific or copyrighted, nor is it verbose.

Edit: Also, looking at their post again, they were trying to probe Muse to see if it knew Mario by using the Wikipedia article. I wouldn't expect that to work because the datasets image generators are trained on only use captions (e.g. "red apple on a wooden table"), not full articles. So, in your defense, all those extra tokens could be adding garbage to the result, as Muse is mostly its own model. Although, in my experience, models tend to just filter out anything they don't know, especially when paired with something they do know; plus, I did like you did and used the Wikipedia article as prompt in some Stable Diffusion models, and it worked just fine.

Yeah, I've been wondering that. Do users have copyright ownership of things generated with Muse? I would think not considering it's still AI, just like anything else.

 

They claim that with the creation there was no copyright of others violated and that's what should (at least theoretically) matter for Steam which bans some AI generated stuff out of copyright precaution.
On the copyright of the output images I couldn't find anything from Unity. However your game as a whole naturally acquires copyright, so it would be really a stretch for someone to extract just the AI generated sprites and use them. Not to mention that that's hardly ever worth it.

Yes and again it's a good thing that this didn't work on Unity's product because it protects you from accidentally using foreign IP in your game.

 

They can claim anything they want. The data is still proprietary and of unknown origins. That same blog post mentions expanding the dataset with Stable Diffusion, which is undergoing several lawsuits. To me this sounds like they've finetuned SD, which is not allowed by Steam because it uses scraped data from the web and for which the legality of using that scraped data is being challenged in court right now.

And you can't prove ownership of Unity's dataset, which is what Steam requires - that you own 100% of the assets you're publishing on their platform. It's a black box, much like anything Unity have been doing recently.

 

Yeah, but I would say it depends on the individual and what they're using it for. For example, if you generate your main character using Muse, you might still want to own the likeness of that character. The wording in the official guidance on works containing AI generated content (more here) makes me think that may not be possible.

For example, does this artist own the likeness of (the face of) this character? (This is an example of acceptable usage of AI generated art on Newgrounds.)

 

You are apparently reading a different text than me if that's the gist you got.

This is the interesting paragraph:

Definitely not a finetuned SD - again proven by the fact that it does not inherently know things like Mario which SD clearly does, as I have shown here.

In that case it will be word vs someone else's word, since Steam cannot prove the opposite either.
I too am curious how this will pan out.

Yeah this is a general topic that laws are not prepared for yet.

Still that aspect shouldn't really be a concern for an indie at least. If your character has become so famous that people are copying it, you have made it big!
And the gaming community usually shuns copycats that aren't clearly fangames anyways.

 

They could've filtered out popular IP just like DALI did. They could've done a million things that are still shady, just like literally every other big AI company in the field. None of them have trained these models ethically, and all of them are undergoing lawsuits. Why is Unity different? Their data is proprietary of unknown origin and that's all folks at Steam care about.

Uh, no. Steam has the final word because it's their platform and their review process and they won't let any more AI content of dubious origin on their platform until they are sure they are not made liable for it. They even ban AI text translations or any reference to AIs in LLM context. None of it is going through the review process if mentioned and/or detected.

 

I don't see Muse to be more reliable than let's say SD in terms of generated content.

Unity doesn't even propry curate Asset Store. Wih tons of assets flips and propieraty concerns models. So Unity claim, that data set for training is any better? Besides removing obvious figures, like Mario. Unity Engine doesnt have enough workforce to currate content. I am yet to see in reality, to believe otherwise.

So for me, Unity Engine properly "curated" content is rather doubtful.

Either way, I don't see issue with generative tools. If someone did use it to make another Mario like game, doesn't matter if is using hand drawing or AI. Developer should knowingly be aware type of content is putting into a game.

It is as old problem as properiaty laws been introduced, before even PC been invented. Mobiles and Steam is flooded wit assets flips. No need for AI to even this to happen. So, I don't see a NEW problem here.

These tools used right way by devs, won't be recognisable as AI generated content.
Devs and artist that knows, won't be scared and stopped by jealousy and crusades lawsuits against AI world progress. All is about AI "taking jobs".
Not to protect ownerships and copy rights.

If they really care, they would take well known issues like hazard elements in children games. But hey, that brings money in long addictive run ...

 

No, because MidJourney may have drawn from copyrighted sources even if at first glance the average person doesn't recognize any of it. For all we know that example character has elements that could be recognizable by someone somewhere on the planet. Maybe that "W" belt buckle is the same as an existing trademarked or copyrighted logo. We just don't know.

 

In copyright law, simple cases take 2 years and complex ones 5 years.
Where a lot of money is involved, powerful firms get hired, this usually means defendant drag things on.
By the time this is sorted out, your PS6 game will be in the 50c bin at Wokemart.

 

Yes, but that's not the only reason: even in the case where it is provably original, according to the official guidance on AI, and copyright law in general, it needs to have been created by a human author and have some amount of creativity (i.e. originality) to have copyright protection. In fact, I think the same rules apply to things generated by any computer program at all (e.g. handwritten procedural generators).

The funny thing is... Midjourney didn't actually make that part, the artist did; Midjourney only made the head, and the artist did the rest. But what you said is still true: there's no guarantee that the belt design doesn't already exist.