Auto Watermark Remover

You have an image with a watermark and you want it gone, and you don't want to hand-paint the mask. Auto Watermark Remover is the whole job in one node: it detects the watermark, builds a mask, and fills the hole, all locally. The name is entirely literal - it calls no API and needs no key.

It's not magic, and the README is refreshingly honest about that. This node works best when the watermark is still visible enough to detect: readable text, corner logos and stock-photo marks, faint semi-transparent overlays with recoverable edges, and repeated symbol watermarks when you can feed it a reference template. If the watermark is a blend that's nearly invisible to your eye, no detector is going to save you.

How it works

Under the hood it's the classic detector-then-inpaint pipeline, which is still the practical open-source frontier for generic watermark removal. Detection happens first - downscaled if your input is huge, then mapped back to a full-resolution mask so big images don't crawl. Then the actual cleanup runs.

The parts that matter, in order:

• Detection. cv2_only uses a contour/text-like-region heuristic and is the fastest starting point. ocr_then_cv2 adds EasyOCR for readable text. ocr_only is for when you specifically want text and no CV2 fallback.
• Hard cases. detail_detector=decomposition recovers faint blended overlays that simple contours miss. detail_detector=symbol_template matches a known logo against a reference image you connect.
• The fill. opencv_telea is the fast first pass for small marks; big_lama is the quality upgrade for thicker or textured areas; symbol_reverse_blend reconstructs a dark semi-transparent symbol from your template; gemini_reverse_alpha is a narrow special case for the Gemini sparkle overlay (yes, that watermark is a real plague - it leaks into models' training data and shows up uninvited); diffusers_sd_inpaint runs a crop-aware SD inpainting pass for semantic cleanup on faces, bodies, and costume detail.

The inputs you'll actually touch

Of the twenty-plus widgets, beginners can ignore most of them. The ones that matter:

• image - your watermarked image, as a batch or a single frame.
• detection_mode - start cv2_only, switch to ocr_then_cv2 when the watermark is readable text.
• region - edges or corners for border watermarks; full_image otherwise. Smaller regions are faster and produce fewer false hits.
• inpaint_method - opencv_telea for speed, big_lama for quality.
• watermark_template - an optional reference crop of the logo or symbol. A tight, clean crop on a similar background beats a noisy crop from the same image every time.

Worth tuning: cv2_sensitivity (raise it for faint overlays, lower it to stop masking real detail), padding and dilate (if watermark edges survive, nudge these up), and symbol_match_threshold (start around 0.45–0.55; raise it if the template keeps latching onto lookalikes).

The node outputs the cleaned image, plus a mask, a mask_preview you can actually look at before trusting it, and detected_text (the OCR'd watermark text and detector messages). Wire image onward to whatever's next; use mask_preview for debugging.

Installing it

ComfyUI Manager is easiest - search for ComfyUI Auto Watermark Mask. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/goodguy1963/comfyui-auto-watermark-mask

Then restart ComfyUI. Manager will install the Python deps, which is where the install actually gets interesting: easyocr, onnxruntime-gpu (on Windows x64 - a chunky install), opencv-python, spandrel, and huggingface_hub. Expect a first-run download or two: EasyOCR pulls its language weights, and Big-LaMa auto-downloads big-lama.pt from Hugging Face into your inpaint model folder (disable that with COMFYUI_AUTO_WATERMARK_MASK_AUTO_DOWNLOAD=0 and drop the file in yourself).

Where people get burned

• symbol_reverse_blend errors out if no template is connected - it requires watermark_template or its legacy alias symbol_reference. Not a bug; it literally cannot run without the reference.
• The Diffusers method isn't installed by default. diffusers_sd_inpaint needs diffusers, transformers, and pillow in your environment, which requirements.txt does not include, and the default model is the ~4GB stable-diffusion-v1-5/stable-diffusion-inpainting. If you pick it without the packages, it fails. Install them only if you actually need semantic cleanup.
• First OCR run feels hung. EasyOCR downloads language weights on first use; it's slow once, fast after.
• Mask eats the subject. cv2_sensitivity too high turns real image detail into "watermark." Check mask_preview before the fill - that output exists precisely so you don't find out after.

It's a niche tool, and the community knows the generic version has limits. For stock-style corner marks and readable text it's genuinely good; for everything else, start with the defaults and work your way up the ladder one method at a time.