Turn NVIDIA PiD into a universal 4x upscaler via FLUX.2 VAE re-encoding

TimeTruth2490 · reddit · 2026-11-06

A Reddit user shares a workaround that turns NVIDIA's Pixel Diffusion Decoder (PiD) into an upscaler for any image.

PiD replaces a model's VAE decoder and redraws images directly in pixels at 4x width/height in 4 steps, but NVIDIA only ships decoders for FLUX.1, FLUX.2, Qwen-Image v1, SD3, SDXL, plus DINOv2/SigLIP.
The trick: re-encode any picture (render from any model, photo, video frame) with the FLUX.2 VAE, then run PiD's FLUX.2 decoder — making it a universal upscaler. Unlike classic upscalers, PiD redraws the image, synthesizing fine detail (skin, hair, texture) rather than sharpening pixels.
Example: 704×384 input becomes 2816×1536 (16x pixels) in about 6 GPU-seconds on Comfy Cloud, clearly beating bicubic 4x.
Ready-made ComfyUI workflows (local and cloud) are published on Hugging Face (lvladikov/PiD-as-pixel-Upscaler); even PiD-incompatible models like Qwen Image 2.1 work with this approach.

More from Multimodal