PiD — Research release¶

Development line: project:pid · thread research-release
Last event: 2026-05-25 · 1 dated since 2026-05-25 · Researched: 2026-09-05 · confidence: high

What it is¶

PiD is a pixel diffusion decoder for generative image pipelines that replaces standard VAE and RAE decoders.

• Latent decoding: decodes latents to pixels with simultaneous 4× upscale, or 8× upscale for Scale-RAE.
• Latent support: works with conventional VAE latents and semantic latents, including DINOv2 and SigLIP.
• Fast inference: distilled checkpoints decode in four steps.

NSCLv1 weights are restricted to non-commercial research or evaluation. Use PiD when you need controlled latent-to-2K/4K decoding on a supported backbone instead of a separate pixel upscaler after the VAE.

Development line¶

• 2026-05-25 — PiD research, code, and model resources were made publicly available. On 2026-05-25, PiD was released with linked research, source code, and Hugging Face model resources. These links establish public technical entry points for the project, though the release evidence does not establish a precise version or full capabilities.

What changed¶

• 2026-05-25 — Published paper, source code, and PiD weights for FLUX, FLUX.2, Z-Image, Z-Image-Turbo, SD3, DINOv2, and SigLIP.
• 2026-05-27 — Added PiD support to ComfyUI.
• 2026-06-02 — Released checkpoints for SDXL, Qwen-Image, and Qwen-Image-2512; cleaned the codebase and enabled torch.compile support.
• 2026-07-09 — Released training code, distilled and undistilled PiD v1.5 2K→4K, and v1.5 checkpoints for FLUX, Z-Image, Z-Image-Turbo, FLUX.2, and Qwen-Image.
• 2026-07-14 — Added optional support for Boogu-Image native generation and PiD decoding of its Flux-style VAE latents.

How to use this¶

From 2026-05-25, practitioners should treat PiD's research page, source repository, and Hugging Face model page as its dated public technical entry points.

1. Clone the repository, create a Python 3.12/CUDA environment with uv sync --frozen or install the dependencies, then run PYTHONPATH=. python verify_env.py. — https://github.com/nv-tlabs/PiD
2. Download only the checkpoints tree: hf download nvidia/PiD --local-dir . --include "checkpoints/*". — https://github.com/nv-tlabs/PiD
3. Use from_ldm with --backbone for prompt-to-image generation; use from_clean to encode an existing image and decode it with PiD. — https://github.com/nv-tlabs/PiD
4. Select --pid_ckpt_type 2k for 2K output; choose 2kto4k_v1pt5 for supported 4K models, or 2kto4k for SD3 and SDXL. — https://github.com/nv-tlabs/PiD

Best practices¶

• Verify the environment with verify_env.py before running inference, and run commands from the repository root with PYTHONPATH=.. — https://github.com/nv-tlabs/PiD
• Do not confuse VAE files with PiD models: PiD_* files are distilled decoder checkpoints, while ae.safetensors, VAE/RAE, and similar files are dependent encoders and decoders. — https://huggingface.co/nvidia/PiD
• Choose v1.5 2K→4K for FLUX, FLUX.2, and Qwen-Image to resolve color and corner-grid artifacts; the 2K checkpoint remains sharper at exactly 2048 px. — https://github.com/nv-tlabs/PiD
• Verify the license before deployment: published weights are restricted to non-commercial research or evaluation. — https://huggingface.co/nvidia/PiD

Superseded by this¶

• 2026-07-09 — Earlier FLUX, FLUX.2, and Qwen-Image 2kto4k checkpoints are deprecated in favor of 2kto4k_v1pt5; older weights moved to checkpoints_deprecated/.
• 2026-07-09 — Advice to use v1 2kto4k for new FLUX, FLUX.2, and Qwen-Image 4K decodes is obsolete; v1 2kto4k remains current for SD3 and SDXL.

Still unknown¶

• The initial README lists dates for subsequent updates, but GitHub does not publish separate versioned releases; exact commit SHAs and timestamps are not established.
• Speed and quality claims come from the authors' own benchmarks; independent reproducible evaluation on a specific workflow was not tested here.

Sources¶

Agent brief¶

• Subject: project:pid, thread research-release, 1 dated events 2026-05-25 → 2026-05-25.
• Practical note: From 2026-05-25, practitioners should treat PiD's research page, source repository, and Hugging Face model page as its dated public technical entry points.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.