NVIDIA Spatial Intelligence Lab proposes PiD, which redesigns the decoding stage in latent diffusion models.

Currently, mainstream text-to-image generation works in the latent space, mapping back to pixels using a VAE decoder. This decoder aims to invert the encoder and reconstruct the image, and does not itself synthesize new details. Moreover, its cost increases significantly at the megapixel scale.

PiD reformulates “latent → pixel” decoding as a conditional pixel diffusion process, unifying decoding and upsampling into a single module.

Decoding a 512×512 latent to 2048×2048 takes less than 1 second on an RTX 5090 (peak memory 13GB), and about 210ms on a GB200. This is about 5.9× faster than the cascaded super-resolution baseline SeedVR2 (211ms vs 1237ms).

Through noise-corrupted latent training and sigma-aware gating, PiD remains robust to partially denoised latents, allowing the base LDM to exit early while PiD handles the remaining high-resolution generation. The method supports both VAE latents and semantic latents (DINOv2 / SigLIP / RAE).

The actual performance and applicability boundaries still need to be verified in more scenarios. However, rethinking the decoder as a diffusion module capable of synthesizing details is a direction worth attention.

https://research.nvidia.com/labs/sil/projects/pid/…

Fast and High-Resolution Latent Decoding with Pixel Diffusion

Source: https://research.nvidia.com/labs/sil/projects/pid/ SD3 PiD Decode

SD3 VAE Decode

VAE Decoder

PiD

DINOv2 PiD Decode

DINOv2 RAE Decode

RAE Decoder

PiD

Z-Image PiD Decode

Z-Image VAE Decode

VAE Decoder

PiD

Flux.2 PiD Decode

Flux.2 VAE Decode

VAE Decoder

PiD

Abstract

Most practical high-resolution text-to-image systems rely on latent diffusion models, where generation is performed in a compact latent space and a decoder maps latents back to pixels. Yet the latent-to-pixel decoder is reconstruction-oriented, optimized to invert the encoder rather than synthesize more details, and becomes increasingly costly at megapixel scale. This drawback calls for a more expressive and efficient decoding paradigm. Motivated by recent progress in scalable pixel-space diffusion, we introducePiD, aPixel diffusionDecoder that reformulates latent decoding as conditional pixel diffusion, unifying decoding and upsampling into one generative module. By denoising directly in high-resolution pixel space,PiDsynthesizes 4× and even 8× upscaled images with low latency. For latent conditioning, a lightweight sigma-aware adapter injects noise-corrupted latents into the pixel diffusion backbone, enablingPiDto decode partially denoised latents and terminate the latent diffusion process early. To further improve efficiency, we distill the model using DMD2, reducing inference to just 4 steps.PiDapplies to both conventional VAE latents and semantic latents (e.g., SigLIP, DINOv2) used in recent RAE-based models.PiDdecodes latents of 512×512 images into 2048×2048 pixels in under 1 second with 13 GB peak memory on a consumer RTX 5090, and as fast as 210 ms on a GB200 GPU, about 6× faster than cascaded diffusion-based super-resolution pipelines with better visual fidelity.

Results

From Latent to Pixels

Select a latent space and move the step slider to compare PiD decoding quality at different early-termination points. Drag thewhite divideron each image to reveal the VAE/RAE decode vs. PiD decode.

4K Decode

Direct latent→4K decoding with PiD. Click any image to launch a side-by-side comparison against the VAE decoder.

Baseline Comparison

Hover over any image to activate the synchronized zoom lens across all six views.

Quantitative Results(Decoding + Upsampling, 5122 → 20482)

End-to-End Decoding Latency (ms) ↓

PiD is up to5.9× fasterthan SeedVR2 (211.2 ms vs 1237.5 ms)

Gemini-3-Flash Judge Rating (%) ↑

% of evaluations where judges preferPiDover each baseline

Method

**Overview of PiD.**PiD unifies latent decoding and upsampling as a single latent-conditioned pixel diffusion model that predicts the target-resolution pixel-space velocity field. Noise-corrupted latent training and sigma-aware gating make the decoder robust to partially denoised latents, enabling early exit from the base LDM while preserving high-resolution output quality.