Nodes/ComfyUI Auto Watermark Mask/Auto Watermark Remover
ComfyUI Node

Auto Watermark Remover

A ComfyUI node in image/cleanup with 27 inputs and 4 outputs.

By goodguy1963·Created 3 months ago·Updated 3 months ago· 0
Auto Watermark Remover
  • image
  • watermark_template
  • symbol_reference
  • image
  • mask
  • mask_preview
  • detected_text
languagesen
detection_modecv2_only
regionfull_image
min_confidence0.25
padding10
dilate5
blur3
cv2_sensitivity0.80
corner_fallbackoff
corner_width_ratio0.12
corner_height_ratio0.08
inpaint_methodopencv_telea
inpaint_modelbig-lama.pt
inpaint_radius5
gpufalse
detail_detectornone
symbol_match_threshold0.45
diffusers_model_idstable-diffusion-v1-5/stable-diffusion-inpainting
diffusers_promptclean image, preserve the original subject and style, remove watermark and text
diffusers_negative_promptwatermark, text, logo, signature, artifact, distortion, deformed anatomy, blurry face
diffusers_strength0.90
diffusers_guidance_scale4.5
diffusers_steps30
diffusers_crop_padding48
Categoryimage/cleanup

Inputs (27)

NameTypeDefaultDescription
imageIMAGE
languagesSTRINGenUsed only when OCR modes are selected.
detection_modeCOMBOcv2_onlyHow to create the watermark mask before cleanup. cv2_only is usually the best starting point for logos and symbols.
regionCOMBOfull_imageLimits detection to likely watermark areas. Smaller regions are faster and reduce false hits.
min_confidenceFLOAT0.250–1Minimum OCR confidence before text is masked. Only matters for OCR modes.
paddingINT100–256Extra pixels around each detected area before cleanup.
dilateINT50–256Expands the final mask. Increase if the watermark edges are still visible after removal.
blurINT30–255Softens the mask edge. Lower values keep the repaired region tighter.
cv2_sensitivityFLOAT0.800–1Sensitivity for non-OCR detection. Raise it for faint overlays, lower it to avoid masking real image detail.
corner_fallbackCOMBOoffAdds a simple mask in one or more corners when the watermark sits near the border and detection is inconsistent.
corner_width_ratioFLOAT0.120.02–0.5Corner fallback width as a fraction of image width.
corner_height_ratioFLOAT0.080.02–0.5Corner fallback height as a fraction of image height.
inpaint_methodCOMBOopencv_teleaCleanup backend. Start with opencv_telea for small marks, big_lama for thicker texture overlap, symbol_reverse_blend when you have a watermark_template for a semi-transparent symbol.
inpaint_modelCOMBObig-lama.ptBig-LaMa checkpoint name. Used only when inpaint_method=big_lama.
inpaint_radiusINT51–64OpenCV or cleanup radius. Small values preserve detail better; larger values can erase thicker marks.
gpuBOOLEANfalseUse GPU for OCR when available. Big-LaMa and Diffusers manage their own device placement.
detail_detectorCOMBOnoneExtra detector for hard masks. Use symbol_template together with watermark_template for known logos or repeated symbol watermarks.
symbol_match_thresholdFLOAT0.450–1Template match threshold for symbol_template and symbol_reverse_blend. Raise it to be stricter.
diffusers_model_idSTRINGstable-diffusion-v1-5/stable-diffusion-inpaintingOptional Hugging Face model id or local folder for diffusers_sd_inpaint.
diffusers_promptSTRINGclean image, preserve the original subject and style, remove watermark and textPrompt used only by diffusers_sd_inpaint. Keep it focused on preserving the original subject while removing the watermark.
diffusers_negative_promptSTRINGwatermark, text, logo, signature, artifact, distortion, deformed anatomy, blurry faceNegative prompt used only by diffusers_sd_inpaint.
diffusers_strengthFLOAT0.900–1How strongly Diffusers redraws the cropped area. Lower values preserve more of the source image.
diffusers_guidance_scaleFLOAT4.51–20Prompt guidance for diffusers_sd_inpaint.
diffusers_stepsINT301–100Inference steps for diffusers_sd_inpaint.
diffusers_crop_paddingINT480–512Extra context around the masked area when using diffusers_sd_inpaint.
watermark_templateoptIMAGEOptional reference image of the watermark to match. Use a tight crop when possible. A clean external template works better than a noisy crop from the same image.
symbol_referenceoptIMAGELegacy alias for watermark_template. You only need one of these inputs connected.

Outputs (4)

NameTypeDescription
imageIMAGE
maskMASK
mask_previewIMAGE
detected_textSTRING