← sources

MOSS Image with akhaliq's 4-step LoRA and Nvidia PiD = Super-Fast High-Res (ComfyUI)

source: reddit (reddit:StableDiffusion)published 2026-10-31 00:25ingested 2026-11-01 02:38

retained text

A few months ago, NVIDIA introduced a groundbreaking image generation module called [PiD (Pixel Diffusion Decoder)](https://huggingface.co/nvidia/PiD) that replaces traditional VAE decoding pipelines. By unifying decoding and upsampling into a single, four-step generative process, it can turn low-resolution latents directly into ultra-sharp 4K images with stunning visual detail in under a second.

Recently I tried it with MOSS Image since its VAE is perfectly compatible out of the box because MOSS actually uses the FLUX.2 VAE, allowing PiD to process those latents flawlessly for high-resolution decoding.

I have created a deep dive on HF where I am providing 2 ComfyUI workflows, one for local and one for cloud, as well as Diffusers Python script if you don't use ComfyUI. I have also generated 100 images using the cloud workflow so you can see for yourself. If you have used PiD in the past, you'll know it may be really good for some images and 'just good' for others, but where it stands out is the performance it brings to the renders - **you get high resolutions 2k+ images very fast** because you **render your images at 1/4 of your target resolution with the model, in this case** [MOSS Image](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) *(also using the recommended* [4 step LoRA by akhaliq](https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps) *for even faster results)*, and [PiD](https://huggingface.co/nvidia/PiD) **turns the small latent straight into an image 4× its width and height**, in **4 more steps**. There is **no VAE decode** and no separate upscaler: **PiD draws the full-size image itself**, guided by the small latent and the prompt.

Full details, review, workflows and examples at my HF: [https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes](https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes)

The 100 examples page, directly: [https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes/blob/main/examples/README.md](https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes/blob/main/examples/README.md)

Update: I've now also added 2 **Refiner workflows** which take on from PiD and further refine the image (with choice between Qwen Image 2.1 and Lightning LoRA, and Krea 2 Turbo using my 3 Step LoRA). Do read the readme in regards to memory usage and how to set ComfyUI to unload models after each use, unless you have plenty of VRAM. One of the workflows if for ComfyUI Cloud and the other for ComfyUI local. This should resolve some of the concerns that PiD makes *plastic* looking results: [https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes#refiner](https://huggingface.co/lvladikov/MOSS-PiD-SuperFast-HighRes#refiner) ; See examples before and after refiner, in the comments section.

document fetched 2026-11-01 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xjw1jd/moss_image_with_akhaliqs_4step_lora_and_nvidia