4 ms·
If I have it generate a “bowl of soup” will I find an identical bowl in some clip art collection somewhere? How much does it deviate from the source images?
by pupppet 5y ago
If I have it generate a “bowl of soup” will I find an identical bowl in some clip art collection somewhere? How much does it deviate from the source images?
- SleekEagle 5y agoYou can try to reverse image search - from what I've seen of other people doing this, the renditions are quite distinct. The diffusion process is ultimately the root of the model's ability to not just copy images. Variational methods truly allow for the learning of a distribution, which is why VAEs can generate new data and AEs can't. Also, practically from a data point of view, the same object can be represented in numerous ways (different artistic styles, different filters, abstract paintings, etc.) and the model has to optimize across all of these samples. What this means is that the model truly is forced to learn the semantic meaning behind a concept and not just rely on specific features. Check out the dropdown under the "Significance of CLIP to DALL-E 2" section in the article
- corysama 5y agoI've played with tech like this for over a year now. You won't find the bowl in the source images. It doesn't evolve the noise into a source image. It slowly nudges the noise into feeling more and more bowl-like. Do that enough, and you get something that feels quite a bit like a bowl. Put it this way: The model file is absurdly smaller than the half billion source images files. If it actually contained the source images, it would be the greatest feat of image compression ever. Instead it only contains the impression left over by the images. A lot closer to a memory than a jpg.