Remove Watermarks from Doubao, Jimeng, Qwen, Tencent Yuanbao & Baidu AI Images: OpenNoMark Local Guide
Remove Watermarks from Doubao, Jimeng, Qwen, Tencent Yuanbao & Baidu AI Images: OpenNoMark Local Guide
The Watermark Trap of China's AI Image Boom
In 2026, China's AI image generation market is exploding — Doubao, Jimeng, Tongyi Wanxiang, Kling, Tencent Yuanbao, Baidu Wenxin. Almost every major platform stamps a watermark in the corner of every generated image. For casual users that is a minor annoyance. For e-commerce asset producers, social media designers, draft mockup creators, and freelancers delivering work to clients, that watermark means the image simply cannot be used as-is.
What makes it worse: there is no unified way to remove these watermarks. Some platforms let paid subscribers turn the watermark off; others offer no clean-export option at all. Online removal tools require uploading image after image — a privacy problem when the material is commercial or unreleased. Is there a solution that is local, open-source, free, and covers every major Chinese AI platform at once?
In July 2026, an open-source project called OpenNoMark appeared on GitHub targeting exactly this pain point. This article breaks down how it works technically, walks through all four ways to use it, and gives an honest verdict on where it fits in the Chinese AI platform watermark landscape.
1. The State of Watermarks on Chinese AI Platforms
Let's first be clear about the problem. Visible watermarks from Chinese AI image/video platforms fall into two broad categories:
- Corner badges: semi-transparent logos or platform names fixed to a corner (usually bottom-right), the most common type
- Text marks: phrases like "AI generated", "for preview only", SAMPLE, or edge-bound URLs and account handles
In its README, OpenNoMark's author states the project ships calibrated fast paths for Gemini, Doubao, Qwen, Jimeng, Kling, Tencent Yuanbao, Baidu, and similar generators — meaning detection has been tuned specifically for the typical positions and shapes of these platforms' watermarks.
For creators, the value of this coverage is obvious: instead of hunting down a separate tool for every platform, one local tool handles the mainstream Chinese platforms plus Gemini.
2. What Is OpenNoMark?
OpenNoMark is a local-first, open-source AI watermark removal tool. Its core selling points:
- Fully local processing: detection and reconstruction run on your own machine; images are never sent to a hosted inference API. This matters for commercial assets, unreleased work, and client deliverables
- Multi-expert watermark localization: "where the watermark is" and "how to reconstruct the missing content" are separated into independent modules, so detection can evolve per generator without rewriting the inpainting stack
- Four delivery forms from one core: the same processing pipeline powers a Web UI, CLI, Python API, and an AI coding-agent Skill, with identical result semantics everywhere
Know its boundary up front: OpenNoMark only handles visible overlays. It does not alter or remove invisible provenance metadata or content credentials (C2PA, EXIF, XMP). If your goal is stripping AI-content provenance information, you'll need complementary tools (see the comparison section).
3. How It Works: Localize, Inpaint, Verify
The technical architecture is the most interesting part of this project, because it explains why it beats one-size-fits-all blur tools.
3.1 Tiered Localization
- Gemini lightweight spatial detector: a dedicated detector for Gemini image watermarks, calibrated on real positive samples and hard negatives — not a fine-tuned foundation model, so it stays fast
- OWLv2 generic detection: runs established corner prompts and generic watermark prompts in separate passes, so new vocabulary can't dilute the calibrated signal
- PP-OCRv5 text fallback: local OCR recognizes phrases like
SAMPLE,PREVIEW,AI GENERATED, and edge-bound URLs/handles, converting the recognized polygon into a tighter mask
3.2 Conservative Fusion
Multiple detectors may propose overlapping regions. The fusion strategy:
- Precise platform masks win over overlapping generic proposals
- Standalone generic regions require strong semantic or matching OCR evidence
- Maximum 4 compact regions; candidate-count and total-area budgets apply
- Tiled, dense, or ambiguous overlays fail closed — the tool refuses to partially erase rather than risk damaging content
3.3 LaMa Inpainting + Visual Validation
- Content-aware LaMa (Large Mask Inpainting) reconstructs the masked region
- After repair, the same visual expert checks the area; residual artifacts trigger one mask-only retry
- Unresolvable evidence is honestly reported as
partial— never silent success
This localize-repair-verify loop, plus inspectable edit metadata (the output reports the detected box and exact feathered mask bounds), lets users distinguish legitimate blending from out-of-area edits — essential for professionals who need to audit deliverable quality.
4. Four Ways to Use It
4.1 Web Workbench (visual review & batches)
For users who prefer not to touch a terminal:
- Requirements: Python 3.10+, uv, Node.js/npm; 16 GB RAM recommended
- Run
start.sh/start.bat, then openhttp://localhost:48292 - Single-image inspection and long-running batches: per-file states (queued, uploading, processing, completed, failed), before/after comparison, per-file retry, ZIP export of all results
- Add more images while others are still processing; finished results are never reprocessed
- English and Simplified Chinese UI
4.2 CLI (folders & automation)
Install without cloning the repo:
uv tool install git+https://github.com/NanmiCoder/OpenNoMark.git opennomark --version
Handles files, paths, directories, or globs (e.g., ./incoming/*.webp), with detection-box/mask debug output, machine-readable JSON responses for scripts and agents, and explicit device selection.
4.3 Python API (application integration)
from opennomark.pipeline import WatermarkRemovalPipeline
pipeline = WatermarkRemovalPipeline()
image, metadata = pipeline.process("image.png", "clean_image.png")
print(metadata["status"], metadata["watermarks_found"])
results = pipeline.process_batch(
["img1.png", "img2.jpg"], output_dir="output/",
callback=lambda index, total, item: print(index, total, item["status"])
)
4.4 Agent Skill (inside coding agents)
Follows the cross-agent skills layout (vercel-labs/skills), working with Codex, Claude Code, Cursor, OpenCode, and other supported agents. The Skill invokes the same versioned CLI (directly or via uvx) — no second image-processing implementation, so results are identical no matter which entry point you use.
5. Models & Hardware Requirements
Model weights are downloaded on first run:
| Model | Purpose | Size |
|---|---|---|
| OWLv2 | Watermark detection | ~500 MB |
| LaMa | Content inpainting | ~196 MB |
| PP-OCRv5 (mobile detector + recognizer) | Generic text fallback | ~45 MB (lazy download) |
Device behavior:
| Environment | Detection | Inpainting |
|---|---|---|
| NVIDIA CUDA | CUDA | CUDA |
| Apple Silicon | MPS when available | CPU fallback |
| CPU-only | CPU | CPU |
Note: LaMa falls back to CPU on MPS (its TorchScript graph contains operations MPS doesn't reliably support), so Apple Silicon users should set expectations accordingly. Batch concurrency is bounded — no model-per-worker.
6. Safety Boundaries & Limitations (Read First)
OpenNoMark is refreshingly honest about its limits. Check these against your needs before adopting it:
- Only small-to-medium, high-confidence visible marks: corner badges and text watermarks are the target
- Not a general object-removal tool: broad or dense tiled candidates are explicitly blocked; don't use it to delete people or large objects from photos
- No invisible watermark or metadata handling: C2PA, EXIF, content credentials are out of scope
- Complex watermarks may report
partial: it will truthfully report residual evidence instead of handing you a "looks clean" result
7. How It Compares
| Option | Platform coverage | Privacy | Cost | Best for |
|---|---|---|---|---|
| OpenNoMark (local, open-source) | Doubao/Jimeng/Qwen/Kling/Yuanbao/Baidu/Gemini etc. | ✅ Fully local | Free | Creators, developers, privacy-conscious users |
| Online watermark removers | One or a few | ❌ Upload required | Subscription/credits | Casual users who don't mind uploading |
| Platform paid plans | Single platform | ✅ | Monthly fee | Heavy users of one platform |
| Manual Photoshop/retouch | Any | ✅ | Time | Low frequency, fine control needed |
For anyone juggling images from multiple Chinese AI platforms who also cares about asset confidentiality, OpenNoMark is one of the few full-coverage local options available.
8. Use Cases & Verdict
Good fit:
- E-commerce/social media: batch-clean Doubao and Jimeng product images before they enter your asset library
- Design freelancing: clean Tongyi and Kling drafts and deliver watermark-free versions to clients
- Internal training/proposals: show AI-generated material without platform logos
- Developers: embed the capability into your content pipeline via the Python API
- AI toolchain enthusiasts: let Claude Code / Codex agents invoke watermark removal directly in a project
Not a fit:
- Stripping C2PA/EXIF provenance metadata (it doesn't do this — pair it with exiftool and friends)
- General object removal with large complex occlusions (not a general inpainting tool)
- Users with zero technical appetite (the Web UI lowers the bar but doesn't eliminate it)
Verdict: OpenNoMark solves the "many platforms, many watermark types, online tools leak my files" problem with a sound multi-expert localization + LaMa inpainting + conservative fusion design. It isn't a flashy online product — it's a solid local tool. If you need to remove AI image watermarks in batch, privately, across platforms, it belongs in your toolkit. The project is iterating fast (29 commits as of late July 2026), and wider platform support is a reasonable thing to expect next.