Salesforce Researchers Introduce XGen-Image-1: A Text-To-Image Latent Diffusion Model

Abstraction: Salesforce latent diffusion text-to-image model reusing pretrained components cheaply

Key points:

• XGen-Image-1 trained on LAION-2B dataset (aesthetic score 4.5+) using TPU v4s at a budget of ~$75K
• Performance matches Stable Diffusion 1.5/2.1 on CLIP Score and FID metrics; exceeds SD on FID
• Key innovation: combining a pretrained VAE with pixel upsamplers enables training at 32x32 resolution while generating 1024x1024 images
• Automated rejection sampling using PickScore evaluation during inference improves output quality
• Latent diffusion approach operates in compressed spatial domain rather than direct pixel space
• Reusing pretrained components (autoencoder, upsamplers) dramatically reduces training cost without sacrificing quality

Connections: Salesforce · Stable Diffusion · Diffusion Models · Text To Image · Generative AI

Source: https://www.marktechpost.com/2023/08/13/salesforce-researchers-introduce-xgen-image-1-a-text-to-image-latent-diffusion-model-trained-to-re-use-several-pretrained-components/