This is a QWEN Image Edit workflow designed to run large AIO QWEN checkpoints (≈28GB) while still generating high-resolution outputs on 12GB VRAM GPUs.

The focus here is:

• Image editing / guided edits

• Very low step counts

• Stable results at low CFG

• Aggressive memory management

• Clean upscale + post polish

If you’ve struggled getting QWEN AIO models to behave on smaller cards, this setup is built specifically to solve that.

Key Features

• Runs QWEN AIO (GGUF or Safetensors) models on 12GB GPUs

• Uses Phroots AIO https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO (Both NSFW & SFW available)

• Tested high sampling resolution

• Works with 1–3 input images for guided edits

• Extremely low step counts (4–6 steps)

• CFG-stable at CFG = 1

• Includes upscaling, resizing, CAS sharpening, and desaturation

• Cleans GPU memory automatically after runs

• LoRA-compatible (QWEN-trained LoRAs supported)

This workflow prioritizes practical generation, not theory — fast previews, predictable edits, and minimal VRAM spikes.

Required Nodes / Extensions

Make sure you have all of these installed:

Core

• ComfyUI (recent)

• ComfyUI-GGUF

• QWEN Image nodes

  • TextEncodeQwenImageEditPlus

  • ClipLoaderGGUF

  • UnetLoaderGGUF

  • VaeGGUF

Upscaling / Image

• was-node-suite

• KJNodes

• ComfyUI Essentials

• 4x_foolhardy_Remacri.pth (upscale model)

Utility / Memory

• easy-use

  • easy clearCacheAll

  • easy cleanGpuUsed

If something errors: double-check GGUF + QWEN nodes first — most issues come from mismatched versions.

Recommended Sampler Settings

These are intentional — higher values usually make QWEN worse, not better.

Sampler: euler_ancestral
Scheduler: beta
Steps: 4–6
CFG: 1.0
Denoise: 1.0
Seed: random or fixed

If you’re coming from SDXL: do not raise CFG. QWEN responds very differently.

LoRA Notes

• QWEN-trained style LoRAs do work

• Load via Model-only LoRA loader

• Suggested strength range:

  • 0.85 → 1.0

• Avoid stacking multiple LoRAs unless you know what you’re doing (VRAM spikes fast)

VRAM & Stability Notes

• Designed to keep peak VRAM under ~12GB

• GGUF models strongly recommended for smaller GPUs

• Cache clearing nodes are intentional — don’t remove them unless you have >24GB VRAM

• If you OOM:

  • Reduce output resolution slightly

  • Close other GPU apps

  • Avoid second diffusion passes (QWEN doesn’t like them)