Nvidia drops PID 1.5 to upscale AI images in seconds

PiD 1.5 is an open-source diffusion decoder from Nvidia that replaces standard VAEs, turning latent representations straight into super-resolved pixels in a single pass.

Most modern image generation pipelines are clunky. Your model generates latents, a VAE decodes them into a standard image, and then a secondary upscaler has to blow it up. PiD throws out that multi-step pipeline by reformulating latent decoding as a conditional pixel-space diffusion model, merging decoding and upscaling into one seamless step.

On a standard consumer GPU, it takes mere seconds to turn a 512x512 latent into a clean 2K image, with direct scaling up to 4K.

What is new in version 1.5

Compared to version 1.0, PiD 1.5 delivers visibly sharper detail and much more accurate color fidelity.

Earlier iterations sometimes suffered from color shifts or subtle grid artifacts tucked away in image corners. Version 1.5 eliminates those corner grid patterns entirely, providing a far cleaner output across different aspect ratios like 16:9 and 9:16.

Nvidia also trained 1.5 on significantly more anime illustrations and small-face data. That means tiny faces in wide shots and delicate line art no longer turn into mush.

Compatibility has widened as well. The 1.5 decoder checkpoints now plug directly into backbones like FLUX, FLUX.2, and Qwen-Image alongside existing setups like Z-Image.

How to get started

The code, weights, and training scripts are fully available on GitHub under nv-tlabs/PiD and on Hugging Face.

If you prefer running inference via Python, you can set up an environment in minutes using uv or pip, provided you already have PyTorch with CUDA and modern diffusers installed. Nvidia provides simple scripts to generate straight from text prompts or upscale clean images.

For visual artists and workflow tinkerers, PiD is already supported in ComfyUI. You can drop it into your existing generation graph to bypass your default VAE entirely.

Keep in mind one small tradeoff mentioned in the documentation: if you only ever need 2048px resolution, the original dedicated 2K checkpoints remain slightly sharper. The 1.5 release is built for the dynamic 2K-to-4K range, delivering better color and cleaner artifacts at higher resolutions.

Why it matters

If you run open-source diffusion models like FLUX locally, PiD 1.5 is an instant upgrade. It cuts out the friction of running dedicated upscaler nodes, saves GPU memory, and slices seconds off every render.

We think Nvidia's approach to replacing standard VAE decoders with a unified generative module is the right direction for local AI generation. It proves you do not need heavier post-processing stacks when your decoder is smart enough to handle super-resolution natively.