ComfyUI custom nodes for remove-ai-watermarks: identify provenance, remove visible and invisible AI watermarks, strip AI metadata, run the complete image pipeline, and erase arbitrary regions inside a ComfyUI graph.

Nodes

Node	What it does
Remove Visible Watermark (RAIW)	Removes every detected registered AI-provenance mark when `mark = auto`, or forces one selected mark. Choose the fill backend and, for `mark = auto`, the detection sensitivity. No GPU is required for detection or the default cv2 fill.
Detect Visible Watermark (RAIW)	Reports per-mark detection confidence on the input image. Outputs a text report, a `detected` boolean, the best confidence, and the detected mark key.
Erase Region (RAIW)	Inpaints whatever a `MASK` covers. Choose cv2 for the dependency-free path or LaMa for a heavier learned fill.
Remove Invisible Watermark / SynthID (RAIW)	Diffusion regeneration through the CUDA-only `qwen-zimage`, `qwen21-zimage`, `sdxl-zimage`, `chroma-zimage`, or `auto` profile.
Identify Provenance (RAIW)	Reads an original file and returns the versioned provenance report, verdict, platform, and confidence.
Strip AI Metadata (RAIW)	Losslessly strips and verifies AI metadata from an original file, then returns the cleaned image and path.
Remove All Watermarks (RAIW)	Runs visible removal, conditional invisible removal, and verified metadata stripping against one original file.

The four pixel nodes operate in-memory on ComfyUI image batches. The three provenance-aware nodes accept a file path because ComfyUI IMAGE tensors do not carry the original C2PA, EXIF, XMP, IPTC, or container data. Their source and output paths must be accessible to the ComfyUI server. A blank output path writes <source>_clean.<ext> beside the source without overwriting it.

Install

Via ComfyUI Manager

Search for "Remove AI Watermarks" in ComfyUI Manager and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/wiltodelta/ComfyUI-remove-ai-watermarks
pip install -r ComfyUI-remove-ai-watermarks/requirements.txt

The package installs remove-ai-watermarks[all,classify-onnx,text-draft], including video tools, HEIF decoding, diffusion, ONNX fill and classification, and OCR proposals. Model weights download on first use.

Notes

ComfyUI image tensors carry no file metadata. The in-memory invisible node therefore uses resolution-adaptive strength for qwen-zimage and the unknown-vendor strength for sdxl-zimage. Remove All Watermarks reads the original file and can use its vendor provenance. Set strength above 0 to override either default.
All five profiles are CUDA-only: qwen-zimage (the default), qwen21-zimage (Qwen Research License, non-commercial research and personal use only), sdxl-zimage, chroma-zimage, and auto (sdxl-zimage for Google, chroma-zimage for Microsoft, qwen-zimage for OpenAI, Meta, and unknown provenance). There is no CPU or MPS path, so the node has no device widget.
The invisible node exposes no steps, guidance_scale, model or device input. Each profile pins its model stack, its per-stage distilled schedule and CFG 1.0, so a widget for any of them could only produce an error inside the run. Seed 0 is the certified default.
sensitivity applies only to mark = auto. An explicit mark skips detection entirely, so there is no borderline decision left to relax or tighten; the node says so in its info output rather than dropping the setting silently.
The inpaint toggle does nothing. Since library 0.16 every visible mark uses localize -> fill, so the switch has no path left to choose. It stays on the node because removing a widget breaks every saved workflow that carries it.
adaptive_polish is three-way. profile default lets the library decide (off for qwen-zimage, whose output already matches the input's detail level; on for sdxl-zimage); on/off override it. A workflow saved when this input was a checkbox still loads, and its stored true/false counts as an override.
Use tile for large inputs and cpu_offload to stream both stacks instead of pinning them in VRAM.
Proprietary invisible-watermark verification is vendor-specific. Use the matching provider oracle. Higher strength removes more but drifts further from the original.
The selected extras install opencv-python-headless. If your ComfyUI install already ships opencv-python, both provide cv2 and coexist; if you hit a cv2 conflict, keep a single OpenCV distribution in the environment.

Release synchronization

The registry package has its own version. Each remove-ai-watermarks release dispatches the synchronization workflow with its exact version. The workflow updates the dependency floor, runs compatibility tests, bumps the node patch version, and publishes only after those tests pass. A daily scheduled run is the recovery path for an interrupted release dispatch. A failed compatibility test blocks publication instead of exposing an incompatible node update.

The package name is the registry id

name in pyproject.toml is the ComfyUI Registry node id, not a local label. Publishing under a different one creates a second listing and orphans every existing install, so it is not something to change casually.

It deliberately matches the PyPI library this node depends on. uv reads that as a self-dependency and refuses to resolve the project, which looks like a naming mistake and is not one: this repository is installed with pip, through requirements.txt and scripts/test.sh in CI, and pip resolves the dependency from PyPI without complaint. There is no uv.lock here and uv run is not part of the workflow.

Editing pyproject.toml on main triggers the publish workflow, so treat any change to this file as a release action rather than an edit.

License

Apache-2.0, matching the upstream library.

Complete operation nodes

The remove-ai-watermarks/operations category exposes every installed library CLI operation with typed controls, plus the ONNX photo classifier and OCR draft proposals. This includes video processing and Seedance, text provenance and regeneration, photo/source classifiers, batch operations, regions, profiles, and quality parameters. Existing image tensor nodes remain available. Operation nodes accept paths on the ComfyUI server and return an output path, report, and exit code. File operations run on each queue, so changed files and deleted outputs are not hidden by workflow caching. They preserve the input; connect output paths to the appropriate image/video/text loader for the next workflow stage. Empty optional controls retain library defaults.

The node package installs all,classify-onnx,text-draft. Use Python 3.12 for the complete backend selection. Image regeneration requires NVIDIA CUDA. Video VAE profiles can use CPU, MPS, or CUDA. Model weights download on first use. OCR results are proposals requiring operator verification, not verified manifest text.