<h3 id="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.">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.</h3>The focus here is:<ul><li>Image editing / guided edits</li><li>Very low step counts</li><li>Stable results at low CFG</li><li>Aggressive memory management</li><li>Clean upscale + post polish</li></ul>If you’ve struggled getting QWEN AIO models to behave on smaller cards, this setup is built specifically to solve that.<hr /><h2 id="key-features">Key Features</h2><ul><li>Runs QWEN AIO (GGUF or Safetensors) models on 12GB GPUs </li><li>Uses Phroots AIO <a target="_blank" rel="ugc" href="https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO">https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO</a> (Both NSFW &amp; SFW available) </li><li>Tested high sampling resolution</li><li>Works with 1–3 input images for guided edits</li><li>Extremely low step counts (4–6 steps)</li><li>CFG-stable at CFG = 1</li><li>Includes upscaling, resizing, CAS sharpening, and desaturation</li><li>Cleans GPU memory automatically after runs</li><li>LoRA-compatible (QWEN-trained LoRAs supported)</li></ul>This workflow prioritizes practical generation, not theory — fast previews, predictable edits, and minimal VRAM spikes.<hr /><h2 id="required-nodes-extensions">Required Nodes / Extensions</h2>Make sure you have all of these installed:<h3 id="core">Core</h3><ul><li>ComfyUI (recent)</li><li>ComfyUI-GGUF</li><li>QWEN Image nodes<ul><li><code>TextEncodeQwenImageEditPlus</code></li><li><code>ClipLoaderGGUF</code></li><li><code>UnetLoaderGGUF</code></li><li><code>VaeGGUF</code></li></ul></li></ul><h3 id="upscaling-image">Upscaling / Image</h3><ul><li>was-node-suite</li><li>KJNodes</li><li>ComfyUI Essentials</li><li>4x_foolhardy_Remacri.pth (upscale model)</li></ul><h3 id="utility-memory">Utility / Memory</h3><ul><li>easy-use<ul><li><code>easy clearCacheAll</code></li><li><code>easy cleanGpuUsed</code></li></ul></li></ul>If something errors: double-check GGUF + QWEN nodes first — most issues come from mismatched versions.<hr /><h2 id="recommended-sampler-settings">Recommended Sampler Settings</h2>These are intentional — higher values usually make QWEN worse, not better.<pre><code>Sampler: euler_ancestral
Scheduler: beta
Steps: 4–6
CFG: 1.0
Denoise: 1.0
Seed: random or fixed
</code></pre>If you’re coming from SDXL: do not raise CFG. QWEN responds very differently.<hr /><h2 id="lora-notes">LoRA Notes</h2><ul><li>QWEN-trained style LoRAs do work</li><li>Load via Model-only LoRA loader</li><li>Suggested strength range:<ul><li><code>0.85 → 1.0</code></li></ul></li><li>Avoid stacking multiple LoRAs unless you know what you’re doing (VRAM spikes fast)</li></ul><hr /><h2 id="vram-and-stability-notes">VRAM &amp; Stability Notes</h2><ul><li>Designed to keep peak VRAM under ~12GB</li><li>GGUF models strongly recommended for smaller GPUs</li><li>Cache clearing nodes are intentional — don’t remove them unless you have &gt;24GB VRAM</li><li>If you OOM:<ul><li>Reduce output resolution slightly</li><li>Close other GPU apps</li><li>Avoid second diffusion passes (QWEN doesn’t like them)</li></ul></li></ul>

Files with SHA256: a0db8a6adf8fa334966c31b4f8b18b9c4a900ccd6115d0d4be76508ffb800671