Chris Donahue
@chrisdonahue.com
Research in generative AI for **human** creativity in music + more.
Assistant professor at CMU CSD, leading the 🎼 G-CLef lab. Part time research scientist at Google DeepMind on the Magenta team (views my own)
Assistant professor at CMU CSD, leading the 🎼 G-CLef lab. Part time research scientist at Google DeepMind on the Magenta team (views my own)
Congrats Kaitlyn and Cornell!!
April 24, 2025 at 4:20 AM
Congrats Kaitlyn and Cornell!!
Also “relative inefficiency of input-space models starts to be economically preferable over the increased engineering complexity of latent-space models”
I wonder about this! If latents shift the scaling laws for generative modeling by an order of magnitude or more, hard to imagine this going away
I wonder about this! If latents shift the scaling laws for generative modeling by an order of magnitude or more, hard to imagine this going away
April 15, 2025 at 11:52 AM
Also “relative inefficiency of input-space models starts to be economically preferable over the increased engineering complexity of latent-space models”
I wonder about this! If latents shift the scaling laws for generative modeling by an order of magnitude or more, hard to imagine this going away
I wonder about this! If latents shift the scaling laws for generative modeling by an order of magnitude or more, hard to imagine this going away
Incredible post. I still don’t have a clear mental model for the need for *both* perceptual and adversarial losses. Seems like they both encourage preservation of certain higher frequency material. Is using both just a hack that works or is there some more fundamental explanation?
April 15, 2025 at 11:50 AM
Incredible post. I still don’t have a clear mental model for the need for *both* perceptual and adversarial losses. Seems like they both encourage preservation of certain higher frequency material. Is using both just a hack that works or is there some more fundamental explanation?