Neural Expedition
Subscribe
Model briefingModel: PiDID: nvidia/pid

Compare native and 4x upscaled generations

This is a useful image pick if you want to see what happens after the first generation is already done. PiD gives you a browser demo for comparing a normal Z-Image result against a higher-resolution pixel-space decode from the same prompt.

PublishedMay 28, 2026
Read time3 min
Tested byNeural Expedition

Field notes

What it does

PiD is a pixel diffusion decoder. In plain terms, it replaces the final step that turns a model's compressed image representation into visible pixels, and it can do that while upscaling the result.

The public demo pairs PiD with Z-Image. You write a prompt, Z-Image creates the initial image internally, and PiD turns the final latent into a larger image. The useful part is the side-by-side slider: you can compare the standard Z-Image decode with the PiD 4x upscale instead of judging the new decoder in isolation.

For a 512px Z-Image run, the demo targets a 2048px result. For a 1024px run, the higher setting targets a 4K output. That makes this less like a generic image generator and more like a test of whether a different final decoding step preserves detail better at inspection size.

How to try it

Start with the Hugging Face Space. Use one prompt where extra resolution should matter: product texture, fabric, detailed fur, small objects on a table, or an architectural scene with repeated lines.

Run the default 512px setting first. When the slider appears, compare edges, texture, small objects, and faces or text if they are present. The question is not only whether the PiD side is larger. The real test is whether the extra pixels hold together when you zoom in.

If that looks promising, try the 1024px setting for a heavier 4K test. For local use, move to the PiD GitHub repo and the nvidia/PiD model page. The released checkpoints include the Z-Image-compatible path, but this is a large GPU workflow rather than a casual laptop setup.

Caveat

PiD is released for non-commercial research and evaluation use. Treat the demo as a way to test image quality and workflow fit, not as a production image service. Local reproduction also needs a strong CUDA setup, especially for the 4K path.

What you can do with it

  • Compare standard and high-resolution decodes from the same image prompt.
  • Test whether product shots, close-ups, and texture-heavy scenes survive upscaling.
  • Generate larger concept images before deciding whether a prompt is worth refining.
  • Inspect where a final image decoder helps, and where it invents brittle detail.
  • Use Z-Image plus PiD as a repeatable benchmark for prompt-to-2K image tests.

Try the demo

View model page