Back to AI Watermark Remover

Open-source models

The two default models are hosted here and load automatically on the tool page. Below are the other compatible open-source models, listed by type with their official addresses; download them yourself — some must be exported to ONNX first. Once downloaded, add it on the tool page under "Add model".

This site only hosts the two default models used by the tool. The files below never pass through this site and are not part of its inference path; use them in your own Python / ComfyUI / ONNX Runtime environment.

The default models (YOLOv8n detector + MI-GAN 512 inpainter) are hosted here and load automatically on the tool page — no download needed.

Compatible detection models

2

Requirements: an Ultralytics YOLOv8 / YOLO11 detector with a fixed 640×640 input, exported with NMS enabled (yolo export model=best.pt format=onnx imgsz=640 nms=True opset=12) so the output is [1,300,6]. Class names may differ; they only affect the labels.

YOLOv8n Watermark DetectionPITAExport it yourself

640×640 · float32 · [1,300,6] · NMSLicense: AGPL-3.0

Model details

YOLOv8n fine-tuned on the PITA dataset for text and logo watermarks. AGPL-3.0; the model card limits it to academic use. The repo ships .pt weights only, so export it to ONNX with the recipe below.

Go to downloadOpen-source project
YOLO11 Watermark DetectorSoliton80Export it yourself

640×640 · [1,300,6] · NMS

Model details

A YOLO11 fine-tune for text, logo, semi-transparent and multilingual watermarks, distributed with a ComfyUI node. Obtain the .pt weights yourself, then export with the recipe below.

Go to download

Compatible inpainting models

2

Requirements: export the whole MI-GAN pipeline with the official script (python -m scripts.create_onnx_pipeline --resolution 256|512). Inputs are uint8 image + mask (255 keep / 0 inpaint), any resolution; output is uint8 RGB.

MI-GAN 256 (Places2)256Export it yourself

uint8 image + mask · dynamic H/WLicense: MIT

Model details

MI-GAN 256 (Places2) for general scenes; smaller and faster than the default 512 build, with slightly softer detail. Convert the official .pt weights with create_onnx_pipeline.py.

Go to download
MI-GAN 256 (FFHQ)256Export it yourself

uint8 image + mask · dynamic H/WLicense: MIT

Model details

MI-GAN 256 (FFHQ) works better on face-centric images. Convert the official .pt weights with create_onnx_pipeline.py.

Go to download

Quick answer

The AI watermark remover ships with two default models — YOLOv8n for watermark detection and MI-GAN 512 for inpainting — both hosted by PureSnap and loaded automatically. This page lists other compatible open-source detectors and inpainting models with their official download addresses; they are not proxied, cached or added to the default pipeline, and some must be exported to ONNX before the browser can run them.

Watermark removal is two tasks: finding the watermark (object detection) and rebuilding what was behind it (inpainting). Each model on this page is one of those two, listed with its licence, expected input size and the export step it needs. A detector and an inpainter must be paired — the browser cannot run one without the other.

Nothing here is downloaded automatically. Each entry links to the project that publishes the model, so the licence that applies is the one stated by that project, not PureSnap's. Imported models run locally in your browser exactly like the defaults: no image is uploaded, and a locally chosen file only lasts for the current session.

Frequently asked questions

Do I need to download a model to use the tool?

No. The two default models are hosted by PureSnap and initialise on first use. Extra models are only for people who want to try a different detector or inpainter.

Why do some models need an ONNX export?

The tool runs models in the browser through ONNX Runtime Web, so weights distributed as PyTorch checkpoints (.pt) are not directly usable. Those entries usually come with an export script in their repository; export to ONNX, then add the resulting file to the tool.

Are third-party models covered by PureSnap's terms?

No. Each model keeps its own licence, shown on its card. Check it before commercial use — some are research-only or require attribution.

Related tools and guides

Last updated 2026-10-26