Stability AI plans to let artists opt out of Stable Diffusion 3 image training

[dill-shower]

On Wednesday, Stability AI announced it would allow artists to remove their work from the training dataset for an upcoming Stable Diffusion 3.0 release. The move comes as an artist advocacy group called Spawning tweeted that Stability AI would honor opt-out requests collected on its Have I Been Trained website. The details of how the plan will be implemented remain incomplete and unclear, however.

As a brief recap, Stable Diffusion, an AI image synthesis model, gained its ability to generate images by "learning" from a large dataset of images scraped from the Internet without consulting any rights holders for permission. Some artists are upset about it because Stable Diffusion generates images that can potentially rival human artists in an unlimited quantity. We've been following the ethical debate since Stable Diffusion's public launch in August 2022.

To understand how the Stable Diffusion 3 opt-out system is supposed to work, we created an account on Have I Been Trained and uploaded an image of the Atari Pong arcade flyer (which we do not own). After the site's search engine found matches in the Large-scale Artificial Intelligence Open Network (LAION) image database, we right-clicked several thumbnails individually and selected "Opt-Out This Image" in a pop-up menu.

Once flagged, we could see the images in a list of images we had marked as opt-out. We didn't encounter any attempt to verify our identity or any legal control over the images we supposedly "opted out."

Other snags: To remove an image from the training, it must already be in the LAION dataset and must be searchable on Have I Been Trained. And there is currently no way to opt out large groups of images or the many copies of the same image that might be in the dataset.

Full news: https://arstechnica.com/information-technology/2022/12/stability-ai-plans-to-let-artists-opt-out-of-stable-diffusion-3-image-training/

[5KilosOfCheese]

Seems sensible. Nothing wrong with this in my opinion. Covers the ass of Stability from legal actions, dodges possible regulators, and keeps copyright holders happy. Other than validating the actual copyright status of the one claiming opt-out still being an issue, I can only see this as a positive thing. Those who some fucking reason are upset about this should just skip out on SD3 and focus on some other base model not made by stability, or just make their own - all the tools are available at huggingface. Honestly more base models is better for this field.

However as we know, nothing is worse than giving collective internet power over you. There will be trolls who will make a bot that will go on to flag every image they can on LAION just out of spite.

[timntorres]

Anyone who knows enough to be able to write a bot, also knows enough not to be spited by Stability AI 😄

[dill-shower]

Those who some fucking reason are upset about this should just skip out on SD3 and focus on some other base model not made by stability, or just make their own - all the tools are available at huggingface. Honestly more base models is better for this field.

To train your model you need powerful equipment which is very expensive. Waifu-diffusion training used 8 A100 80GB GPUs and 192GB of RAM . The average user of stable-diffusion-webui does not have such hardware. And for SD2.1 training, 32 x 8 x A100 GPUs were used. Therefore the recommendation "if you don't like it train your model" is difficult to implement

[5KilosOfCheese]

Those who some fucking reason are upset about this should just skip out on SD3 and focus on some other base model not made by stability, or just make their own - all the tools are available at huggingface. Honestly more base models is better for this field.

To train your model you need powerful equipment which is very expensive. Waifu-diffusion training used 8 A100 80GB GPUs and 192GB of RAM . The average user of stable-diffusion-webui does not have such hardware. And for SD2.1 training, 32 x 8 x A100 GPUs were used. Therefore the recommendation "if you don't like it train your model" is difficult to implement

Checking my local supplier. To get 1 A100 80Gb, it is about 10.000€. Other related hardware to get one A100 system running costs about 15.000€. So the equivalent of a somewhat used few years old car. I know people who get motorcycles, do skydiving, have cars as a hobby, who spend more than that in an year or two. Getting that machinery is totally doable. However, you don't need to buy that stuff, there are plenty of machine time services available for hire in hourly or project basis. If you have few maters or even a small community - this is totally doable. If you are someone who thinks everything must be free and don't want to invest anything to this - then yeah... You are shit out of luck; in life in general.

But we have seen the machine resource requirements drop dramatically in just few months. Give it few more and I'm sure you can do the base model training with some funky system with consumer hardware like RTX 40xx cards or even less. Currently it would seem that VRAM requirements are in the real of 12-24G for most basic training and finetuning operations. After which the costs are basically just electricity, which is something that not even Stability or OpenAI can get away from.

This stuff is quite literally in the realm of a dedicated average user, if they want to take on the project of doing this. However; I find it irritating that there are people hating and spiting a company that is on their own path and direction; which is giving them stuff for fucking free to use and play with. While also being deadset on not spending a penny on this stuff.

On side note. I know people who have for fun and hobby spent lot of money to run dedicated servers for games so they, their friends, and their community can play. A small community of people can totally and easilly take on a task like this; just like a dedicated one person can. All the code and resources are available online for free.

However once you start an undertaking like this, at some point - especially if you are a group/community - you will need to deal with legal things relating to this. Such as copyright. Whether it is the dataset, using the dataset, origin of the media in the dataset, or the outoput generations, or whatever; you can't get away from this, especially if you want to use the model or training for commercial purposes. This is exactly what Stability is trying to get ahead of; Find a way for copyright holders to opt-out and then you won't have to deal with issues from them.

Besides since SD came out to public use I been saying at every turn I can; we need one base model that has been trained on completely copyright free, and material which has been given with consent by the license holder for inclusion in the model. If we had a base model like this, you could skirt so many issues. Once again, this is totally doable: There are plenty of copyright free image archives, you can easilly set up a system to ask people for submissions; the hardware is totally reasonably priced for a community or even small scale commercial operation's budget; the code is all available.

I mean like pete sake. To get a machine to do this trainig with is nothig compared to if you'd want to start operating as a welder, machine shop, cafe. Hell in my late teens and early 20's I played in non-profit orchestra's that managed to pay us tickets to tour in Europe and USA - and occasionally get stipends for instruments. Instruments which mind you cost about what an any A-series card costs retail. And they managed to pay for rehearsal space, sheet music and licenses, and storage space. And they had been going on for years doing this. Bigger things have been achieved with less.

About Waifu and NAI. The developers of these models choise to use copyrighted content; they chose to use booru sites for images, just like Stability chose to use LAION. All misery and worry about copyright related issues are things they chose to deal with. Now Stability wants to find a way to choose not to deal with this. Which if they manage to figure out a good and functional system for, other model developers can copy and use; this can and will only benefit AI training on commercial and community level.

Besides... The old models are not going anywhere. I know there are people dead set on sticking to SD 1.x models and keep going on with them. There is nothing wrong with that. But if you aren't willing to put the slightest bit of money, time or effort in to this; then you don't get to complain that those who do, aren't doing what you want them to do.

[kalkal11]

Those who some fucking reason are upset about this should just skip out on SD3 and focus on some other base model not made by stability, or just make their own - all the tools are available at huggingface. Honestly more base models is better for this field.

To train your model you need powerful equipment which is very expensive. Waifu-diffusion training used 8 A100 80GB GPUs and 192GB of RAM . The average user of stable-diffusion-webui does not have such hardware. And for SD2.1 training, 32 x 8 x A100 GPUs were used. Therefore the recommendation "if you don't like it train your model" is difficult to implement

Training a whole model is very demanding, training what you actually care about into a model is very doable on a 12gb GPU

I fintuned 90 concepts into a model in just 14 hours, just so long as the base model understands all required basic concepts, this is a practical solution.

[dill-shower]

I fintuned 90 concepts into a model in just 14 hours, just so long as the base model understands all required basic concepts, this is a practical solution.

Dreambooth or hypernetwork?

[learnin-tap]

您好，已收到您的邮件。