5 ms·

Technically correct on the point that it has not been settled in courts yet. For eg stable diffusion I actually dove a bit deeper on this: The LAION dataset i

Technically correct on the point that it has not been settled in courts yet. For eg stable diffusion I actually dove a bit deeper on this: The LAION dataset is -in itself- almost indisputably legal: It contains URLs of images, and tags describing those images[1]. To be sure: _It does not contain actual images_. People have made this mistake before, and LAION quite correctly reply that they don't have them. [2] Stable diffusion's models are much smaller [3] than the LAION dataset which they use as their source. Again, the models do not contain images. This is pretty suggestive that SD models should not be considered derivative works. Of course sometimes funny things happen in court, so we can't be sure until a court actually decides. I don't currently run my own GPT or LLAMA, but those should have similar properties. (in fact, OpenAI appears to claim so.) [1] https://laion.ai/blog/laion-5b/ https://laion.ai/blog/laion-5b/ (if you look for the files in any of the variations, you'll see it's in the order of terrabytes) [2] https://www.vice.com/en/article/pkapb7/a-photographer-tried-to-get-his-photos-removed-from-an-ai-dataset-he-got-an-invoice-instead https://www.vice.com/en/article/pkapb7/a-photographer-tried-... [3] https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 https://huggingface.co/stabilityai/stable-diffusion-xl-base-... : see under "files and versions", it's sd_xl_base_1.0.safetensors. File size is 6.9 GB

Certhas 3y ago
Not sure that the size of the model is relevant. If you take a PNG of Mickey Mouse, and create a similar vector graphic that is a fraction of the size, that would still be derivative.
satyrnein 3y ago
Thanks for the links! To me (merely an amateur observer), it is not perfectly obvious that the LAION dataset is legal, because I don't think "we don't have your original images anymore, we only kept the results of applying an algorithm to them" is a universal defense. If the algorithm is "resize to 640×480" you are probably still infringing, but if it's "calculate md5 checksum" you're probably not. I don't know where on the spectrum LLM training will fall, legally!
- Kim_Bruning 3y ago
  Heh, it does in fact calculate the MD5 checksum! In total, the fields in the LAION dataset are: image_path caption NSFW similarity LICENSE url key status error_message width height original width original height exif md5 https://huggingface.co/datasets/laion/laion2b-multi-vit-l-14-embeddings https://huggingface.co/datasets/laion/laion2b-multi-vit-l-14... Note that LAION is a metadata dataset, it's not an image generation model itself. (But it has been used to train one)

Certhas 3y ago

Not sure that the size of the model is relevant. If you take a PNG of Mickey Mouse, and create a similar vector graphic that is a fraction of the size, that would still be derivative.

satyrnein 3y ago

Thanks for the links! To me (merely an amateur observer), it is not perfectly obvious that the LAION dataset is legal, because I don't think "we don't have your original images anymore, we only kept the results of applying an algorithm to them" is a universal defense. If the algorithm is "resize to 640×480" you are probably still infringing, but if it's "calculate md5 checksum" you're probably not. I don't know where on the spectrum LLM training will fall, legally!

Kim_Bruning 3y ago
Heh, it does in fact calculate the MD5 checksum! In total, the fields in the LAION dataset are: image_path caption NSFW similarity LICENSE url key status error_message width height original width original height exif md5 https://huggingface.co/datasets/laion/laion2b-multi-vit-l-14-embeddings https://huggingface.co/datasets/laion/laion2b-multi-vit-l-14... Note that LAION is a metadata dataset, it's not an image generation model itself. (But it has been used to train one)