June 2024 Summaries

AuraSR is a newly released open-source upsampler model with 600 million parameters, derived from the GigaGAN paper, designed to upscale low-resolution images to four times their original resolution, with the possibility of repeated application. It stands out for its ability to efficiently upscale images generated by text-to-image models without any constraints on resolution or upscaling factor. By leveraging the speed of GANs, AuraSR offers a significant advantage over diffusion and autoregressive models, producing 1024px images in just 0.25 seconds. This release is credited to the contributions from lucidrains and is part of fal.ai's initiative, which is also seeking individuals interested in training open-source models.

Fal has announced several major updates, including the availability of Stable Diffusion 3 (SD3), an advanced text-to-image model developed by stability.ai, which promises enhanced image quality, improved prompt adherence, and superior text generation capabilities. Users can explore this model through fal's dashboard and model playground. Additionally, fal has introduced ComfyUI Workflows, a serverless runtime allowing users to run workflows efficiently without the need for machine setup, with the ability to scale any workflow into an API using fal's serverless engine. Fal has also revamped its dashboard, featuring enhanced analytics, more detailed billing and usage information, and convenient keyboard shortcuts for quicker navigation, inviting users to provide feedback and explore the new features.