5 ms·
DALL-E so far has no way to create consistent characters across multiple pictures. However good news if this discriminator approach https://blog.salesforceaires
by freemint 4y ago
DALL-E so far has no way to create consistent characters across multiple pictures. However good news if this discriminator approach https://blog.salesforceairesearch.com/gedi/ https://blog.salesforceairesearch.com/gedi/ caries over to images such constraints could be imposed. This is just a handful of follow up publications away.
- ducktective 4y agoThanks!
- freemint 4y agoActually i need to correct myself. Something along these lines. The mechanism of DALL-E is a bit different but a discriminator approach should work anyway although maybe some magic might be needed to ensure the spatial invariance of the discriminator.
- Gigachad 4y agoThere was an anime art generator posted here recently which could do this. If I understood correctly, they manually went through all the parameters and worked out which ones correlated with things like hair color, face type and built a web UI which could fix certain known meaningful parameters and allow the user to design a character and then the final step showed the same character in a bunch of different poses.
- gwern 4y agoYes, you reverse-engineer the latent. This is a very old approach in GANs by this point, going back probably to like 2015. (The canonical examples were making faces smile or adding sunglasses.) If you want to see it done for a lot of different models, check out Artbreeder: that's how all of the editing attributes are done. Diffusion models don't really have a latent but the CLIP embedding would serve the same function. The problem is, OA would have to implement it themselves. There's no way you can implement it as a user with the current interface. (This is also true of alternative methods like gradient ascent or GEDI etc.)