PiD: Pixel Diffusion Decoder

Model Overview

PiD (Pixel Diffusion Decoder) is a generative model and decoding paradigm developed by the NVIDIA Spatial Intelligence Lab (SIL) (Yifan Lu, Qi Wu, Sanja Fidler, et al.).

It reformulates latent-to-pixel decoding as a conditional pixel-space diffusion process, replacing traditional Variational Autoencoder (VAE) decoders in latent diffusion pipelines. By unifying decoding and spatial super-resolution (4× to 8× upsampling) into high-resolution pixel space (up to 2048×2048 / 4K), PiD achieves sharp texture synthesis in just 4 steps via DMD2 distillation.

---

Key Features

• Unified Decoding & Upsampling: Combines latent-to-pixel decoding with 4×/8× spatial super-resolution into a single generative pixel-diffusion step.
• Sigma-Aware Adapter: Injects noise-corrupted latents into the pixel diffusion backbone, allowing latent diffusion to terminate early.
• Fast 4-Step Distilled Inference: Achieves 2048×2048 resolution output in 211ms on NVIDIA GB200 or <1s on RTX 5090.
• Universal Latent Compatibility: Functions as a drop-in replacement for standard VAE, SigLIP, and DINOv2 latents across FLUX, SD3, and Z-Image.
• High Detail Synthesis: Eliminates the soft, blurry artifacts of traditional reconstruction VAE decoders.

---

Verified Project Links

• Project Website: https://research.nvidia.com/labs/sil/projects/pid/
• arXiv Paper: https://arxiv.org/abs/2605.23902
• GitHub Repository: https://github.com/nv-tlabs/PiD
• Hugging Face Model: https://huggingface.co/nvidia/PiD

---

Performance Benchmarks

• Decoding Latency (512×512 → 2048×2048): 211.2 ms (GB200) / <1.0 s (RTX 5090).
• Inference Steps: 4 steps via DMD2 Distillation.
• 5.9× acceleration over standard cascaded super-resolution pipelines (SeedVR2).