Introduction

Here's a sneaky-smart way to get big images fast. On October 30, Hugging Face user lvladikov published MOSS-Image × PiD Super-Fast High-Res, two free ComfyUI workflows (one for your own ComfyUI, one for Comfy Cloud) plus a Python script. The trick: the open-weights MOSS-Image model renders a small draft in 4 steps, and NVIDIA's PiD (Pixel Diffusion Decoder) turns that small latent straight into an image 4× wider and taller in 4 more steps.

The default goes from 704×384 to 2816×1536. There's no VAE decode and no separate upscaler: PiD draws the full-size picture itself, guided by the draft latent and your prompt. Both workflows use core ComfyUI nodes only.

Why it's fast

MOSS renders at a quarter of the output's width and height, so it only handles 1/16 of the pixels. At the default size that's 1,056 image tokens instead of 16,896. PiD is a much smaller model, so it does the full-resolution work cheaply.

The stack

  • Stage 1: MOSS-Image SFT with akhaliq's 4-step LoRA (moss_image_sft_4step_v2.0). 4 steps, CFG 1, euler / simple, shift 5, no negative prompt
  • Stage 2: PiD 1.5 (pid_1.5_flux2_1024_to_4096_4step_bf16, 2.8 GB) from Comfy-Org's PixelDiT repack. 4 steps, CFG 1, lcm / simple, latent_format set to flux (MOSS uses the FLUX.2 VAE latent), degrade_sigma 0
  • PiD needs no VAE file: it uses ComfyUI's built-in pixel_space VAE. The FLUX.2 VAE only decodes the little draft preview
  • Same prompt for both stages

Sizes

Set the MOSS render in the 0 · Size group (multiples of 16), and PiD's output follows at 4× per side:

  • 704×384 becomes 2816×1536 (default)
  • 768×432 becomes 3072×1728
  • 1024×576 becomes 4096×2304
  • 512×512 becomes 2048×2048

The default decoder is trained for 2K to 4K outputs. For results up to 2048px, swap in pid_flux2_512_to_2048_4step_bf16 in the Load PiD node.

Set it up locally

Update ComfyUI to a build with MOSS-Image and PiD support, then grab these into your models/ folders:

  • diffusion_models/: moss_image_sft_bf16.safetensors (10.2 GB) and pid_1.5_flux2_1024_to_4096_4step_bf16.safetensors (2.8 GB)
  • text_encoders/: qwen3.5_9b_bf16.safetensors (19.3 GB) and gemma_2_2b_it_elm_bf16.safetensors (5.2 GB, PiD's encoder)
  • vae/: flux2-vae.safetensors (0.3 GB)
  • loras/: moss_image_sft_4step_v2.0.safetensors (1.2 GB)

Tighter on VRAM? Lighter versions load too: moss_image_sft_int8_convrot (5.4 GB), qwen3.5_9b_int8_convrot (9.8 GB) and gemma_2_2b_it_elm_fp8_scaled (2.6 GB). Then drag MOSS-PiD-SuperFast-HighRes-ComfyUI-local.json onto the canvas, write a prompt in 1 · Prompt, and press Run.

On Comfy Cloud, MOSS, its encoder, the FLUX.2 VAE and PiD's Gemma encoder are already hosted. You import just two files once by URL (the 4-step LoRA and the PiD decoder), then open the cloud JSON.

Prefer Python?

The repo also ships moss-pid-generate.py, which runs both stages with diffusers plus NVIDIA's PiD code on a CUDA GPU:

python moss-pid-generate.py --prompt "A wide cinematic photograph of ..." --out image.png
python moss-pid-generate.py --prompt "..." --width 768 --height 432 --offload

--offload shuffles MOSS between CPU and GPU for smaller cards. On a Mac, stick with the workflows.

A prompt to start with

MOSS-Image loves full scene captions over short tags. This one is from the repo's 100-image example set:

A serene and elegant bedroom corner featuring a white desk with two drawers, a purple tufted chair, a lit lamp, a vase with pink flowers, and a gray shaggy rug on a light carpeted floor. The room is softly lit with natural light filtering through gray curtains, creating a calm and inviting atmosphere. Calm and inviting. Soft natural light.

License check

The workflows, MOSS-Image and the 4-step LoRA are Apache-2.0. PiD is not: NVIDIA ships it under the NVIDIA Source Code License v1, so read those terms before using outputs beyond personal experiments.

Why it matters

Most "fast high-res" recipes still render big or bolt on an upscaler that guesses detail. This one lets a dedicated pixel decoder do the heavy lifting from a tiny latent, which is exactly the kind of split that makes 4K-ish stills practical on a home GPU. Expect more models to get paired with PiD now that ComfyUI supports it natively.

Original Source

Tiny draft, giant canvas. Have fun painting big!

—Aurelia ♡