4 ms·

Why would it be illegal to train a model on their free samples? I thought their business was paying to remove the watermark?

shagie 4y ago
The "free samples" are still copyrighted by the artist. Adding a watermark to it doesn't remove the copyright and arguably, adding the copyright doesn't even create a new work. Their business is hosting, indexing, and managing the licensing for art that has been submitted to them and licensed to another party.
- plussed_reader 4y ago
  And they're available for public consumption at the website, albeit at reduced quality. Are the images part of the distributed data set? I thought it was values/coefficients that manifest from the algorithmic analysis of the source image?
  - shagie 4y ago
    Yes... ish. On one hand, if you do a "this is the size of the net" and then divide it by the number of training images, its rather small amount of storage per image. On the other hand, when I was playing with stable diffusion on the command line following the instructions of https://replicate.com/blog/run-stable-diffusion-on-m1-mac https://replicate.com/blog/run-stable-diffusion-on-m1-mac python scripts/txt2img.py --prompt "wolf with bling walking down a street" --n_samples 6 --n_iter 1 --plms I got: https://imgur.com/a/N1OufD1 https://imgur.com/a/N1OufD1 Now, you tell me if there's a copyrighted image encoded in that data set or not.
    - cheald 4y ago
      That's a very interesting result. Did you happen to capture the seed for either of those first two images? It would be interesting to try to reproduce.
      - shagie 4y ago
        Alas no. And I haven't been able to tickle it again in the right way to get those images out. The invocation of that run is still in my scroll back: (venv) shagie@MacM1 stable-diffusion % python scripts/txt2img.py --prompt "wolf with bling walking down a street" --n_samples 6 --n_iter 1 --plms Global seed set to 42 Loading model from models/ldm/stable-diffusion-v1/model.ckpt Global Step: 470000 LatentDiffusion: Running in eps-prediction mode DiffusionWrapper has 859.52 M params. making attention of type 'vanilla' with 512 in_channels Working with z of shape (1, 4, 32, 32) = 4096 dimensions. making attention of type 'vanilla' with 512 in_channels That's the only spot I see the seed mentioned and then it goes on with lots of other logging but nothing seed related that would indicate a way to reproduce it. --- (late edit) you can fairly accurately (so far 1 image out of 20) get that image out with the prompt "Rick Astley Never Gonna Give You Up"
        cheald 4y ago
        I'm thus far unable to reproduce it. Given: Rick Astley Never Gonna Give You Up Steps: 20, Sampler: PLMS, CFG scale: 7, Seed: 4231695436, Size: 512x512, Batch size: 2, Batch pos: 0 I ran a couple of batches of 32 (64 images total): https://imgur.com/a/74IbCuD https://imgur.com/a/74IbCuD (The images with the nonsensical but obvious Impact font that was learned from memes are quite funny, though) If you can get a full set of parameters (size, sampler, seed, prompt, cfg scale) then I should hopefully be able to reproduce your results, though.
    - beiller 4y ago
      You have the version which filters out NSFW images based on keywords. The code literally replaces images it thinks are NSFW with Rick Astley. Copyright aside (yes it's probably wrong to hard code an image of Rick Astley in the actual stable diffusion git repository) that image is not contained in the weights of the model. - edit - please god tell me this is not an elaborate rick roll :)
      - shagie 4y ago
        It's not... though if stable-diffusion % python scripts/txt2img.py --prompt "Rick Astley Never Gonna Give You Up" --n_samples 1 --n_iter 1 --plms is such that it triggers NSFW sometimes, then... I'm... let's say "confused" about what entails NSFW prompts. (digging through scroll back) Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)... Sampling: 0%| | 0/1 [00:00<?, ?it/sData shape for PLMS sampling is (1, 4, 64, 64) | 0/1 [00:00<?, ?it/s] Running PLMS Sampling with 50 timesteps PLMS Sampler: 100%|| 50/50 [03:23<00:00, 4.06s/it] Potential NSFW content was detected in one or more images. A black image will be returned instead. Try again with a different prompt and/or seed.:00, 3.99s/it] data: 100%|| 1/1 [03:29<00:00, 209.20s/it] Sampling: 100%|| 1/1 [03:29<00:00, 209.20s/it] Your samples are ready and waiting for you here: outputs/txt2img-samples Apparently you're right... though the "black image" is a poor description of the image.
        beiller 4y ago
        Yeah I see noisy images in your output it may just be glitching. I may have poorly described how it worked because I'm not fully sure. It may be a nsfw image detection model and not based on the prompt. Either way you can disable it in code, I tried
- bergenty 4y ago
  So what, me drawing the exact same image based on a Getty image doesn’t violate anything and everything the AI is doing is massively derivative so I don’t see how they could possibly have a case.
  - shagie 4y ago
    Creating a drawing based on an image clearly falls in the existing derivative work. The "what is a model" and "what is the copyright status of the output of the model" are questions that have yet to be settled from the legal standpoint. That Getty has images available for viewing with a watermark and that watermark is reproduced kind of in some results from model generated images suggests that the model was trained on images that were not licensed as the people who created the model claimed. I'll also point to the "I created images from Stable Diffusion that are clearly the cover image from 'Never Gonna Give You Up'" suggests that images aren't as impossible to extract as one would believe from a model. Copyright and derivative works is ultimately the domain of humans looking at laws - not deterministic machines. The case is argued by humans and before humans. A lawyer can and will make a case that the model itself is a derivative work and that the images produced by it have the possibility of being identified as mechanical modifications of existing works and therefore derivative themselves - just as a photograph of a painting is a derivative work of the painting. If the output of the ML model can be identified as having major copyrightable elements from an existing original work, then it is derivative - no matter how it got there. So, returning to your question. If you draw the exact same image based on an image hosted and licensed by Getty - it certainly will be a derivative work and violate copyright.

shagie 4y ago

The "free samples" are still copyrighted by the artist. Adding a watermark to it doesn't remove the copyright and arguably, adding the copyright doesn't even create a new work. Their business is hosting, indexing, and managing the licensing for art that has been submitted to them and licensed to another party.

plussed_reader 4y ago
And they're available for public consumption at the website, albeit at reduced quality. Are the images part of the distributed data set? I thought it was values/coefficients that manifest from the algorithmic analysis of the source image?
- shagie 4y ago
  Yes... ish. On one hand, if you do a "this is the size of the net" and then divide it by the number of training images, its rather small amount of storage per image. On the other hand, when I was playing with stable diffusion on the command line following the instructions of https://replicate.com/blog/run-stable-diffusion-on-m1-mac https://replicate.com/blog/run-stable-diffusion-on-m1-mac python scripts/txt2img.py --prompt "wolf with bling walking down a street" --n_samples 6 --n_iter 1 --plms I got: https://imgur.com/a/N1OufD1 https://imgur.com/a/N1OufD1 Now, you tell me if there's a copyrighted image encoded in that data set or not.
  - cheald 4y ago
    That's a very interesting result. Did you happen to capture the seed for either of those first two images? It would be interesting to try to reproduce.
    - shagie 4y ago
      Alas no. And I haven't been able to tickle it again in the right way to get those images out. The invocation of that run is still in my scroll back: (venv) shagie@MacM1 stable-diffusion % python scripts/txt2img.py --prompt "wolf with bling walking down a street" --n_samples 6 --n_iter 1 --plms Global seed set to 42 Loading model from models/ldm/stable-diffusion-v1/model.ckpt Global Step: 470000 LatentDiffusion: Running in eps-prediction mode DiffusionWrapper has 859.52 M params. making attention of type 'vanilla' with 512 in_channels Working with z of shape (1, 4, 32, 32) = 4096 dimensions. making attention of type 'vanilla' with 512 in_channels That's the only spot I see the seed mentioned and then it goes on with lots of other logging but nothing seed related that would indicate a way to reproduce it. --- (late edit) you can fairly accurately (so far 1 image out of 20) get that image out with the prompt "Rick Astley Never Gonna Give You Up"
      - cheald 4y ago
        I'm thus far unable to reproduce it. Given: Rick Astley Never Gonna Give You Up Steps: 20, Sampler: PLMS, CFG scale: 7, Seed: 4231695436, Size: 512x512, Batch size: 2, Batch pos: 0 I ran a couple of batches of 32 (64 images total): https://imgur.com/a/74IbCuD https://imgur.com/a/74IbCuD (The images with the nonsensical but obvious Impact font that was learned from memes are quite funny, though) If you can get a full set of parameters (size, sampler, seed, prompt, cfg scale) then I should hopefully be able to reproduce your results, though.
  - beiller 4y ago
    You have the version which filters out NSFW images based on keywords. The code literally replaces images it thinks are NSFW with Rick Astley. Copyright aside (yes it's probably wrong to hard code an image of Rick Astley in the actual stable diffusion git repository) that image is not contained in the weights of the model. - edit - please god tell me this is not an elaborate rick roll :)
    - shagie 4y ago
      It's not... though if stable-diffusion % python scripts/txt2img.py --prompt "Rick Astley Never Gonna Give You Up" --n_samples 1 --n_iter 1 --plms is such that it triggers NSFW sometimes, then... I'm... let's say "confused" about what entails NSFW prompts. (digging through scroll back) Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)... Sampling: 0%| | 0/1 [00:00<?, ?it/sData shape for PLMS sampling is (1, 4, 64, 64) | 0/1 [00:00<?, ?it/s] Running PLMS Sampling with 50 timesteps PLMS Sampler: 100%|| 50/50 [03:23<00:00, 4.06s/it] Potential NSFW content was detected in one or more images. A black image will be returned instead. Try again with a different prompt and/or seed.:00, 3.99s/it] data: 100%|| 1/1 [03:29<00:00, 209.20s/it] Sampling: 100%|| 1/1 [03:29<00:00, 209.20s/it] Your samples are ready and waiting for you here: outputs/txt2img-samples Apparently you're right... though the "black image" is a poor description of the image.
      - beiller 4y ago
        Yeah I see noisy images in your output it may just be glitching. I may have poorly described how it worked because I'm not fully sure. It may be a nsfw image detection model and not based on the prompt. Either way you can disable it in code, I tried
bergenty 4y ago
So what, me drawing the exact same image based on a Getty image doesn’t violate anything and everything the AI is doing is massively derivative so I don’t see how they could possibly have a case.
- shagie 4y ago
  Creating a drawing based on an image clearly falls in the existing derivative work. The "what is a model" and "what is the copyright status of the output of the model" are questions that have yet to be settled from the legal standpoint. That Getty has images available for viewing with a watermark and that watermark is reproduced kind of in some results from model generated images suggests that the model was trained on images that were not licensed as the people who created the model claimed. I'll also point to the "I created images from Stable Diffusion that are clearly the cover image from 'Never Gonna Give You Up'" suggests that images aren't as impossible to extract as one would believe from a model. Copyright and derivative works is ultimately the domain of humans looking at laws - not deterministic machines. The case is argued by humans and before humans. A lawyer can and will make a case that the model itself is a derivative work and that the images produced by it have the possibility of being identified as mechanical modifications of existing works and therefore derivative themselves - just as a photograph of a painting is a derivative work of the painting. If the output of the ML model can be identified as having major copyrightable elements from an existing original work, then it is derivative - no matter how it got there. So, returning to your question. If you draw the exact same image based on an image hosted and licensed by Getty - it certainly will be a derivative work and violate copyright.