<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dec89a05-65ce-4af2-b138-56863be06f5d/original=true/dec89a05-65ce-4af2-b138-56863be06f5d.jpeg" />Removes watermarks, logos, website addresses and text overlays with Qwen Image 2.1, then pastes only the cleaned areas back into your original photo. Everything the watermark didn't cover keeps its original pixels and the original resolution.About 15 seconds for a typical 1–2 MP photo on an RTX 5090. Core ComfyUI nodes only, no custom nodes.<h3 id="features">Features</h3><ul><li>One click: load a photo, press Run.</li><li>Automatic change mask: the edit is compared with your photo and only the areas that changed are replaced. In my tests 79–98% of pixels stayed identical to the original.</li><li>Colour-matched paste-back: cleaned areas are matched to the photo's colours, so they don't leave darker or lighter shapes of the watermark.</li><li>Original resolution: the model works at about 1 MP, but your photo is never downscaled.</li><li>Painted mask (optional): paint over anything that's left in the Mask Editor; it is added to the automatic mask.</li><li>Only my painted mask: a switch that replaces only what you painted, for photos where the model also removes text that belongs to the scene (a sign, a print on clothing).</li><li>Previews: a before/after slider, the replaced areas, and the full model output.</li></ul><h3 id="how-it-works">How it works</h3><ol><li>Qwen Image 2.1 redraws the photo with the instruction to remove the watermarks and keep everything else the same (12 steps, about 1 MP).</li><li>The edit is compared with your photo at the same size. The blurred difference is thresholded, grown and softened into a mask.</li><li>The edit's colours are histogram-matched to a watermark-free first pass, so the pasted areas blend in.</li><li>The matched edit is scaled to your photo's size and composited into the original through the mask.</li></ol><h3 id="requirements">Requirements</h3><ul><li>A recent ComfyUI with Qwen Image 2.1 support. Tested on ComfyUI 0.37.0 with frontend 1.53.6.</li><li>No custom nodes.</li><li>Models, all from <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1">Comfy-Org/Qwen-Image-2.1</a>. The loader nodes carry these links, so ComfyUI offers to download missing files.<ul><li><code>diffusion_models/</code> <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/diffusion_models/qwen_image_2.1_int8_convrot.safetensors">qwen_image_2.1_int8_convrot.safetensors</a> (7.3 GB)</li><li><code>text_encoders/</code> <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_w4a8.safetensors">qwen3vl_8b_w4a8.safetensors</a> (6.3 GB)</li><li><code>vae/</code> <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors">qwen_image_2.1_vae_bf16.safetensors</a> (0.7 GB)</li></ul></li><li>Tested on an RTX 5090 (32 GB).</li></ul><h3 id="how-to-use">How to use</h3><ol><li>Load your photo into the Photo node and press Run.</li><li>The result is saved to <code>output/Qwen21/watermark/</code>. Replaced areas shows what was changed.</li><li>If something is left, right-click the photo, choose Open in MaskEditor, paint over it and run again.</li></ol><h3 id="settings">Settings</h3><ul><li>Generate, steps: 12. Anything from 8 to 20 works; fewer steps stay closer to your photo.</li><li>Edit instruction, resolution: 1024 (about 1 MP). For photos over about 2 MP use 1536: sharper replaced areas, about 3× slower.</li><li>Sensitivity: 0.12. Lower catches fainter watermarks but replaces more of the photo.</li><li>Grow: 6 px at the edit's size around detected changes (4 px for painted areas). Soften mask edges: radius 6.</li><li>Seed is fixed. If you change only the mask settings, just the fast mask part runs again, because the edit is cached.</li></ul><h3 id="tips-and-limits">Tips and limits</h3><ul><li>Don't add a turbo or lightning LoRA. The 4-step turbo LoRA I tested shifts the picture by about 15 px, which breaks the paste-back. The full model at 12 steps doesn't shift.</li><li>On large photos the replaced areas are softer than the rest, because they come from a 1 MP edit. Raise the resolution for those.</li><li>A watermark over a near-white area changes the photo very little and can be partly missed. Paint it, or lower Sensitivity.</li><li>The instruction is plain text in Edit instruction, so you can adapt it, for example to a date stamp.</li><li>Qwen Image 2.1's VAE returns RGBA. The Edit as RGB node drops the alpha channel; keep it if you change the workflow, or the colour matching breaks.</li></ul><h3 id="tested-on">Tested on</h3>Three photos with synthetic watermarks: a tiled semi-transparent pattern with a site address (3840×2400), an opaque logo (1920×1200), and a framed logo with a handle (1080×1920). All watermarks were removed, 79–98% of pixels stayed untouched, and each run took 14–23 s including saving.<hr />Please use it on images you own or have permission to edit. Removing a watermark from someone else's work to repost it can infringe their copyright.

Ultra Watermark Remover (Qwen Image 2.1)

First release.<ul><li>Qwen Image 2.1 (int8) edit at 12 steps, about 1 MP.</li><li>Automatic change mask, optional painted mask and an "Only my painted mask" switch.</li><li>Histogram colour matching against a watermark-free first pass.</li><li>Tested on ComfyUI 0.37.0 (frontend 1.53.6) with an RTX 5090.</li></ul>

Files with SHA256: 8868fdc0ec051e68668f6978f061d9301d30d36b1e0a5dbdcfa1d417282586a8