NVIDIA pixel-diffusion decoder that turns latents into 2K–4K images, replacing VAE decode plus a separate upscaler.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
PiD (Pixel diffusion Decoder) is an NVIDIA research model that reformulates latent decoding as conditional pixel diffusion. It merges latent-to-pixel decoding and upsampling into one generative step, producing 4x and up to 8x upscaled images directly from latents of diffusion models. A sigma-aware adapter lets it decode partially denoised latents, so the upstream latent diffusion can stop early. A DMD2-distilled variant runs in 4 steps.
According to the project page, PiD decodes 512×512 latents into 2048×2048 pixels in under 1 second with 13 GB peak memory on an RTX 5090, and in about 210 ms on a GB200. It reports up to 5.9x lower latency than SeedVR2 (211.2 ms vs 1237.5 ms) with judges preferring its output. It works with conventional VAE latents and with semantic latents (SigLIP, DINOv2) used by RAE-based models. Checkpoints in ComfyUI cover Flux.1, Flux.2, Qwen Image, SD3 and SDXL latent spaces, at 512→2048 and 1024→4096.
ComfyUI supports PiD and the related PixelDiT 1.3B pixel-space text-to-image model natively. Repackaged weights and workflow templates are hosted by Comfy-Org on Hugging Face. Creators describe using it as an upscaler for existing images, and note that the PixelDiT text-to-image mode at 1K is weaker. The weights use an NVIDIA license (listed as "nsclv1", a non-commercial research license by name), so commercial use should be checked against the license. The GitHub URL provided is the ComfyUI integration (GPL-3.0), not the model's own code.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source Nvidia pixel-space diffusion model that upscales 1K images to 4K (or 512 to 2K) in seconds, running in ComfyUI. Creator shows upscaling demos, claims it is up to 5.9x faster than SeedVR2 with better detail, and walks through install and three workflows including a weak 1K text-to-image mode.
Pixel diffusion decoder that upscales images to 2K+ with strong detail, replacing the latent decode plus separate upscaler step. Upscales 512 to 2K in under a second, about 6x faster than SeedVR2; works with Flux 2, Z-Image, SD3.
Closest neighbours by meaning, not just by category.
Model · NVIDIA
NVIDIA's open-source image upscaler, v1.5, for latent, Flux and Z-Image outputs, targeting fast 2K results on consumer GPUs.
Model · Nvidia
NVIDIA research model that turns a single image into a queryable 3D Gaussian scene with semantics and object instances.
Model · Nvidia
Nvidia research model that understands and generates images and video directly in pixel space, with no VAE or vision encoder.
Model
Pixel-space text-to-image diffusion model transferred from latent diffusion models, without a VAE, supporting 4K output.
Model
Research model that unifies 3D scene generation and reconstruction in a single pixel-space diffusion model.
Model · NVIDIA
Open few-step diffusion refiner from NVIDIA that regenerates low-resolution H3 video drafts at 768P to 4K, 5–15 seconds.
