UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs — mentioned by ollin

Thu, 01 Feb 2024 01:34:07 GMT

some points that stood out to me: 1. they made a lot of careful tweaks to the unet network architecture - it seems like they ran many different ablations here ("In total, our endeavor consumes approximately 512 TPUs spanning 30 days"). 2. the model distillation is based on previous UFOGen work from the same team https://arxiv.org/abs/2311.09257 (hence the UFO graphic in the diffusion-gan diagram) 3. they train their own 8-channel latent encoder / decoder ("VAE") from scratch (similar to Meta's Emu paper) instead of using the SD VAEs like many other papers do 4. they use an internal dataset of…

Paper and discussion history. A mention is not an endorsement.

UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs — mentioned by GaggiX

Fri, 15 Dec 2023 21:42:07 GMT

Running SD v1.5 using one single step is faster than sampling with GigaGAN and you can achieve better coherency in my opinion: https://arxiv.org/abs/2311.09257 Then it does make more sense to run this on a raspberry pi.

Paper and discussion history. A mention is not an endorsement.

UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs · HackerPapers discussions

UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs — mentioned by ollin

UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs — mentioned by GaggiX