Back in June, we shared that while our goal continues to be shipping as many games as possible on Steam, we needed some time to learn about the fast-moving and legally murky space of AI technology, especially given Steam's worldwide reach. Today, after spending the last few months learning more about this space and talking with game developers, we are making changes to how we handle games that use AI technology. This will enable us to release the vast majority of games that use it.
Key points:
new questions for devs submitting their games (about pre- and live-generated AI content, guardrails preventing generation of anything illegal)
disclaimers on game’s store page
new system for reporting illegal content straight from in-game overlay
I remember there being a lot of uncertainty about the legality of what and how can('t) be used in training models (especially when used for commercial purposes) - has that been settled in any way? I think there was also a case of not being able to copyright AI generated content due to lack of human authorship (I’d have to look for an article on this one as it’s been a while) - this obviously won’t be a problem if generated assets are used as a base to be worked upon.
In the United States, the Authors Guild v. Google case established that Google’s use of copyrighted material in its books search constituted fair use. Most people agree this will apply to generative models as well since the nature of the use is highly transformative.
I recommend reading this article by Kit Walsh, a senior staff attorney at the EFF from April last year if you haven’t already. The EFF is a digital rights group who recently won a historic case: border guards now need a warrant to search your phone.
Works involving the use of AI are copyrightable, but just like everything else, it depends. It’s also important to remember the Copyright Office guidance isn’t law. Their guidance reflects only the office’s interpretation based on its experience, it isn’t binding in the courts or other parties. Guidance from the office is not a substitute for legal advice, and it does not create any rights or obligations for anyone. They are the lowest rung on the ladder for deciding what law means.
As for illegal content - Valve mentioned it in regards to live-generated stuff. I assume they’re worried about possibility of plagiarism and things going against their ToS, which is why they ask about guardrails used in such systems. On a more general note, there were also cases of AI articles coming up with fake stories with accusations of criminal behavior involving real people - this probably won’t be a problem with AI usage in games (I hope anyway) but it’s another sensitive topic devs using such tools have to keep in mind.
I agree live generated stuff could get developers in trouble. With pre-generated assets you can make sure ahead of time everything is above board, but that’s not really possible when you have users influencing what content appears in your game. If they were going to ban anything, the original ban should have been limited to just this.
Thanks for the links, that’s exactly why I wasn’t sure where things stand currently. While I am familiar with EFF, I wasn’t aware of that article so it was an interesting read.
The one I kind of remembered (even though only partially) was the Reuters article, which contains this quote I was referring to:
The office reiterated Wednesday that copyright protection depends on the amount of human creativity involved, and that the most popular AI systems likely do not create copyrightable work.
It’s obviously a bit more complicated than how I mentioned it initially so I’m glad I could read it again.
The original ban was always meant to be temporary as far as I understand, Valve simply wanted some time to decide rather than make a rash decision (it’s easier to open the floodgates than it is to clean up after the fact). I’m sure things will change in the future as AI tools become more and more common anyway.
The one I kind of remembered (even though only partially) was the Reuters article, which contains this quote I was referring to:
The office reiterated Wednesday that copyright protection depends on the amount of human creativity involved, and that the most popular AI systems likely do not create copyrightable work.
This was likely in reference to Midjourney, which was the system in question in its ruling. Midjourney, even for its time had very rudimentary user controls way behind the open standards that likely didn’t impress the registrar.
There’s also a spectrum of involvement depending on what tool you’re using. I know with web based interfaces don’t allow for a lot of freedom due to wanting to keep users from generating things outside their terms of use, but with open source models based on Stable Diffusion you can get a lot more involved and get a lot more freedom. We’re in a completely different world from March 2023 as far as generative tools go.
Take a look at the difference between a Midjourney prompt and a Stable Diffusion prompt.
To break down a bit of what’s going on here, I’d like to explain some of the elements found here.
sarasf is the token for the LoRA of the character in this image, and `` is the character LoRA for Sarah from Shining Force II. LoRA are like supplementary models you use on top of a base model to capture a style or concept, like a patch. Some LoRA don’t have activation tokens, and some with them can be used without their token to get different results.
The .07 in `` refers to the strength at which the weights from the LoRA are applied to the output. Lowering the number causes the concept to manifest weaker in the output. You can blend styles this way with just the base model or multiple LoRA at the same time at different strengths. You can even take a monochrome LoRA and take the weight into the negative to get some crazy colors.
The Negative Prompt is where you include things you don’t want in your image. (worst quality, low quality:1.4), here have their attention set to 1.4, attention is sort of like weight, but for tokens. LoRA bring their own weights to add onto the model, whereas attention on tokens works completely inside the weights they’re given. In this negative prompt FastNegativeV2 is an embedding known as a Textual Inversion. It’s sort of like a crystallized collection of tokens that tell the model something precise you want without having to enter the tokens yourself or mess around with the attention manually. Embeddings you put in the negative prompt are known as Negative Embeddings.
In the next part, Steps stands for how many steps you want the model to take to solve the starting noise into an image. More steps take longer.
VAE is the name of the Variational Autoencoder used in this generation. The VAE is responsible for working with the weights to make each image unique. A mismatch of VAE and model can yield blurry and desaturated images, so some models opt to have their VAE baked in,
Size is the dimensions in pixels the image will be generated at.
Seed is the number representation of the starting noise for the image. You need this to be able to reproduce a specific image.
Model is the name of the model used, and Sampler is the name of the algorithm that solves the noise into an image. There are a few different samplers, each with their own trade-offs for speed, quality, and memory usage.
CFG is basically how close you want the model to follow your prompt. Some models can’t handle high CFG values and flip out, giving over-exposed or nonsense output.
Hires steps represents the amount of steps you want to take on the second pass to upscale the output. This is necessary to get higher resolution images without visual artifacts. Hires upscaler is the name of the model that was used during the upscaling step, and again there are a ton of those with their own trade-offs and use cases.
After ADetailer are the parameters for Adetailer, an extension that does a post-process pass to fix things like broken anatomy, faces, and hands. We’ll just leave it at that because I don’t feel like explaining all the different settings found there.
In the United States, the Authors Guild v. Google case established that Google’s use of copyrighted material in its books search constituted fair use. Most people agree this will apply to generative models as well since the nature of the use is highly transformative.
I recommend reading this article by Kit Walsh, a senior staff attorney at the EFF from April last year if you haven’t already. The EFF is a digital rights group who recently won a historic case: border guards now need a warrant to search your phone.
Works involving the use of AI are copyrightable, but just like everything else, it depends. It’s also important to remember the Copyright Office guidance isn’t law. Their guidance reflects only the office’s interpretation based on its experience, it isn’t binding in the courts or other parties. Guidance from the office is not a substitute for legal advice, and it does not create any rights or obligations for anyone. They are the lowest rung on the ladder for deciding what law means.
I agree live generated stuff could get developers in trouble. With pre-generated assets you can make sure ahead of time everything is above board, but that’s not really possible when you have users influencing what content appears in your game. If they were going to ban anything, the original ban should have been limited to just this.
Thanks for the links, that’s exactly why I wasn’t sure where things stand currently. While I am familiar with EFF, I wasn’t aware of that article so it was an interesting read.
The one I kind of remembered (even though only partially) was the Reuters article, which contains this quote I was referring to:
It’s obviously a bit more complicated than how I mentioned it initially so I’m glad I could read it again.
The original ban was always meant to be temporary as far as I understand, Valve simply wanted some time to decide rather than make a rash decision (it’s easier to open the floodgates than it is to clean up after the fact). I’m sure things will change in the future as AI tools become more and more common anyway.
This was likely in reference to Midjourney, which was the system in question in its ruling. Midjourney, even for its time had very rudimentary user controls way behind the open standards that likely didn’t impress the registrar.
There’s also a spectrum of involvement depending on what tool you’re using. I know with web based interfaces don’t allow for a lot of freedom due to wanting to keep users from generating things outside their terms of use, but with open source models based on Stable Diffusion you can get a lot more involved and get a lot more freedom. We’re in a completely different world from March 2023 as far as generative tools go.
Take a look at the difference between a Midjourney prompt and a Stable Diffusion prompt.
Midjourney:
a 80s hollywood sci-fi movie poster of a gigantic lemming attacking a city, with the title "Attack of the Lemmy!!" --ar 3:5 --v 6.0Stable Diffusion:
sarasf, 1girl, solo, robe, long sleeves, white footwear, smile, wide sleeves, closed mouth, blush, looking at viewer, sitting, tree stump, forest, tree, sky, traditional media, 1990s \(style\),Negative prompt: (worst quality, low quality:1.4), FastNegativeV2Steps: 21, VAE: kl-f8-anime2.ckpt, Size: 512x768, Seed: 2303584416, Model: Based64mix-V3-Pruned, Version: v1.6.0, Sampler: DPM++ 2M Karras, VAE hash: df3c506e51, CFG scale: 6, Clip skip: 2, Model hash: 98a1428d4c, Hires steps: 16, "sarasf_V2-10: 1ca692d73fb1", Hires upscale: 2, Hires upscaler: 4x_foolhardy_Remacri, "FastNegativeV2: a7465e7cc2a2",ADetailer model: face_yolov8n.pt, ADetailer version: 23.11.1, Denoising strength: 0.38, ADetailer mask blur: 4, ADetailer model 2nd: Eyes.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur 2nd: 4, ADetailer confidence 2nd: 0.3, ADetailer inpaint padding: 32, ADetailer dilate erode 2nd: 4, ADetailer denoising strength: 0.42, ADetailer inpaint only masked: True, ADetailer inpaint padding 2nd: 32, ADetailer denoising strength 2nd: 0.43, ADetailer inpaint only masked 2nd: TrueTo break down a bit of what’s going on here, I’d like to explain some of the elements found here.
sarasfis the token for the LoRA of the character in this image, and `` is the character LoRA for Sarah from Shining Force II. LoRA are like supplementary models you use on top of a base model to capture a style or concept, like a patch. Some LoRA don’t have activation tokens, and some with them can be used without their token to get different results.The .07 in `` refers to the strength at which the weights from the LoRA are applied to the output. Lowering the number causes the concept to manifest weaker in the output. You can blend styles this way with just the base model or multiple LoRA at the same time at different strengths. You can even take a monochrome LoRA and take the weight into the negative to get some crazy colors.
The Negative Prompt is where you include things you don’t want in your image.
(worst quality, low quality:1.4),here have their attention set to 1.4, attention is sort of like weight, but for tokens. LoRA bring their own weights to add onto the model, whereas attention on tokens works completely inside the weights they’re given. In this negative promptFastNegativeV2is an embedding known as a Textual Inversion. It’s sort of like a crystallized collection of tokens that tell the model something precise you want without having to enter the tokens yourself or mess around with the attention manually. Embeddings you put in the negative prompt are known as Negative Embeddings.In the next part,
Stepsstands for how many steps you want the model to take to solve the starting noise into an image. More steps take longer.VAEis the name of the Variational Autoencoder used in this generation. The VAE is responsible for working with the weights to make each image unique. A mismatch of VAE and model can yield blurry and desaturated images, so some models opt to have their VAE baked in,Sizeis the dimensions in pixels the image will be generated at.Seedis the number representation of the starting noise for the image. You need this to be able to reproduce a specific image.Modelis the name of the model used, andSampleris the name of the algorithm that solves the noise into an image. There are a few different samplers, each with their own trade-offs for speed, quality, and memory usage.CFGis basically how close you want the model to follow your prompt. Some models can’t handle high CFG values and flip out, giving over-exposed or nonsense output.Hires stepsrepresents the amount of steps you want to take on the second pass to upscale the output. This is necessary to get higher resolution images without visual artifacts.Hires upscaleris the name of the model that was used during the upscaling step, and again there are a ton of those with their own trade-offs and use cases.After
ADetailerare the parameters for Adetailer, an extension that does a post-process pass to fix things like broken anatomy, faces, and hands. We’ll just leave it at that because I don’t feel like explaining all the different settings found there.https://youtu.be/-JQDtzSaAuA?t=97
https://youtu.be/1d_jns4W1cM
https://www.youtube.com/watch?v=HtbEuERXSqk
Damn, that’s a good chunk of info! Thanks for taking the time to go into details on how things work.
You’re very welcome. My head still hurts.
Is this your comfy workflow, or from someone else?
Someone else.