NVIDIA PiD upscaler script brings FLUX2VAE to 24GB GPUs

A developer has released a Python script that simplifies NVIDIA's PiD (Progressive Image Diffusion) model for image upscaling, specifically using the FLUX2VAE variant. The script, hosted on GitHub under the repository '3090_shorts', strips away all training-related code from the original NVIDIA repository, leaving only the core inference components powered by torch and transformers. This makes it accessible for developers who want to run high-quality upscaling without diving into complex ComfyUI workflows. However, the VRAM requirements are steep: 24GB for 1024px outputs and 32GB for anything above that, effectively limiting it to high-end consumer and workstation GPUs like the RTX 3090/4090 or professional A-series cards.

The developer notes that the model itself performs well, but flags a significant concern: NVIDIA has changed the license for PiD, as indicated by a screenshot linked in the Reddit post. This creates uncertainty around commercial use and redistribution. The stripped-down script could democratize access to NVIDIA's upscaling technology for enthusiasts and researchers, but the licensing shift may push users toward alternatives like ESRGAN or Real-ESRGAN. Despite the VRAM barrier, the project highlights a trend of companies open-sourcing models with evolving restrictions, forcing the community to navigate legal gray areas.