zlacker

They've got a copy of a figure from the original diffusion paper, showing a diffusion process on a spiral dataset. They seem to completely misunderstand it. The figure does not show image diffusion, rather it shows a diffusion process in which each data item is a 2D point. The figure is showing diffusion on an entire dataset and demonstrating that it can approximately reconstruct the spiral-shaped distribution.

I'm surprised they couldn't find someone with even a rudimentary understanding of diffusion models to review this.

>>Imnimo+(OP)
Wow, not only do they get this wrong, it’s the core example they use to demonstrate copying.

>>grandm+v2
Yeah, and their next figure isn't any better. They show a latent space interpolation figure from DDPM, and they seem to think this is how Diffusion models produce a "collage" (as they describe the process). Of course, this figure has nothing to do with how image generation is actually performed. It's just an experiment for the purpose of the paper to demonstrate that the latent space is structured.

In fact, this only works because the source images are given as input to the forward process - thus, the details being interpolated are from the inputs not from the model. If you look at Appendix Figure 9 from the same paper (https://arxiv.org/pdf/2006.11239.pdf) it is clear what's going on. Only when you take a smaller number of diffusing (q) steps can you successfully interpolate. When you take a large number of diffusing steps (top row of figure 9), all of the information from the input images is lost, and the "interpolations" are now just novel samples.

It's very hard for me to find a reason to include Figure 8 but not Figure 9 in their lawsuit that isn't either a complete lack of understanding, or intentional deception.

>>Imnimo+(OP)
I don’t know if it’s that they lack the understanding or they aim to avoid that understanding because it largely invalidates the lawsuit. In general taking abstract concepts and styles from things is acceptable, not protected by copyright, and crucial to creative exploration. The burden of proof is on those claiming SD is not doing this.

>>Imnimo+(OP)
Oh my god, I'm glad someone else noticed this because it was driving me nuts. Deeply troubling that their case hinges on such a fundamental misunderstanding of the technology.

I bet a proper analysis of that toy experiment would conclude that none of the original data points are perfectly recovered: Only the underlying distribution / manifold is recovered, which really doesn't lend well to their argument at all.