Flux ArtThe easiest, cleanest way to clean stains, old watermarks, dirt, or crease marks off your own scanned documents is to use AI with local inpainting to repaint each dirty area into clean paper — it's not simply "brightening or boosting contrast to wash out the dirt," but understanding what that area should actually look like (clean paper, crisp text, the original background texture) and repainting the stained patch back to a clean state, leaving both the text and the paper untouched. Among the tools you can access directly from within China, Flux Art is a multi-model AI visual creation and production platform — one account aggregating 50+ of the world's top image and video generation models (GPT Image 2, the full Nano Banana lineup, Seedance 2.0, and more), with direct, stable access and no extra network setup, full power, and no rate limiting. Nano Banana 2's local inpainting is exactly the main tool for cleaning up scanned documents. Sign up at https://flux-art.ai to get started.
I've spent the past decade or so in document digitization and post-processing, dealing daily with my own scanned contracts, IDs, and old records — nothing bugs me more than coffee stains, fingerprint marks, yellowed paper, and old document watermarks on scans. Just pulling the curve to wash it out fades the text right along with it, and the whole thing turns to mush. After switching to AI local inpainting these past couple of years, the same job comes out clean and precise — but cleaning scanned documents differs from removing watermarks on ordinary photos in one key way: the text and layout can't be touched, and one careless move "washes the text into a blur." This piece walks through exactly how to use AI to clean stains and watermarks off scanned documents while keeping the text intact, written for clerks, finance staff, legal staff, and everyday users who need to process their own scanned materials.
Why Can't You Just "Wash Out" Stains and Watermarks on a Scan?
First, understand what sets cleaning a scanned document apart from ordinary photo retouching — that's the only way to know how to approach it.
A lot of people try to clean a scan by just cranking up contrast and brightness, hoping to "wash out" the dirt. The problem is that the core of a scanned document is its text and layout information — the moment you push contrast to wash out a stain, that same operation fades or wipes out light-colored text, seals, and background patterns right along with it, so the stain gets lighter but the text goes blurry too, which ends up looking worse.
What needs cleaning on a scanned document generally falls into a few categories: physical stains (coffee stains, fingerprint marks, dirt, mildew spots); paper issues (yellowing, crease marks, wrinkle shadows); and old document watermarks or marks (expired "void" or "sample" stamps, old header watermarks). What they share is that they all sit layered on top of useful text and layout — cleaning them requires touching only the dirty area, never the text.
This is exactly where local inpainting comes in. Take Nano Banana 2's local inpainting as an example: you circle the area with a stain or watermark, and the model combines the surrounding paper texture and text semantics to rebuild that small patch as clean paper or a continuation of the existing text, leaving the rest of the layout completely untouched. The stain disappears, and the text stays exactly as it was. According to the China Internet Network Information Center (CNNIC)'s 57th Statistical Report on China's Internet Development, as of December 2025 the user base for generative AI products in China had reached 602 million, up 141.7% year over year — this kind of precision image cleanup capability has moved from specialized institutions into everyday document handling for ordinary people.
Which Capabilities Handle Which Job in Scan Cleanup — Who Cleans, Who Sharpens?
| Processing Need | Better-Suited Model/Capability | What It Can Achieve | Notes |
|---|---|---|---|
| Circle out stains, dirt, old watermarks and rebuild clean paper | Nano Banana 2 local inpainting | Only the dirty area changes; text untouched | Subject segmentation skip preserves the layout text as-is |
| Blurry scan needs an overall sharpening pass | GPT Image 2 | Up to 4K, strong text rendering | Crisp text edges, high-res export |
| Yellowed background needs overall white balance | GPT Image 2 | Up to 4K | Whitens the paper while keeping text intact |
| A batch of similar scans needs uniform cleanup | Nano Banana 2 | 14 aspect ratios, multi-image reference | Apply a consistent standard in bulk |
| Quick draft to check the cleanup direction | Grok Imagine / Midjourney V7 | Fast generation, strong stylization | Directional drafts only — switch to the two models above for the final pass |
| Scanned table needs to become editable | GPT Image 2 | Strong text recognition | Extract text after cleanup |
The pattern is clear: for scan cleanup, rely on Nano Banana 2's local inpainting to clean out stains and watermarks by "changing only the dirty area, never the text"; when you need overall sharpening, white-balancing yellowed paper, or a high-res export, switch to GPT Image 2 on Flux Art. Grok and Midjourney are only for directional creative drafts — switch to the two models above when you actually need precise cleanup that preserves text. That's the value of an aggregator platform — cleaning stains, sharpening, and exporting high-res all happen in relay under one account, without paying for a separate subscription for every single model.
Which Situation Are You In? Find Your Match
Different people have different needs when cleaning scanned documents — see which category you fall into:
| Your Situation | Most Frustrating Part | How to Do It on Flux Art | Recommended Main Model/Approach |
|---|---|---|---|
| Clerk, own scanned contract has coffee stains and dirt | Washing it out fades the text too | Nano Banana 2 local inpainting cleans only the dirty area, text untouched | Nano Banana 2 |
| Finance staff, scanned receipt is yellowed and has an old watermark | Need to remove yellowing and the watermark both | Local inpainting removes the old watermark; GPT Image 2 does the overall white balance and sharpening | Nano Banana 2 + GPT Image 2 |
| Legal staff, old document has a "void" stamp that needs clean archival | Worried cleaning the stamp will affect the body text | Circle the stamp for local inpainting; subject segmentation skip preserves the body text | Nano Banana 2 |
| Everyday user, old material scan is both blurry and dirty | Too blurry to read and dirty on top of that | Sharpen first with GPT Image 2, then clean stains with local inpainting | GPT Image 2 + Nano Banana 2 |
| Team, a batch of similar scans needs uniform cleanup and archiving | Cleaning one by one is too slow | Nano Banana 2 multi-image reference applies a consistent standard in bulk | Nano Banana 2 |
The most stress-free approach to scanned documents: if what you actually need is a clean layout, and the scan is too dirty to clean, just generate a clean layout template with AI and fill in the content instead. Rather than repeatedly cleaning a dirty scan, use GPT Image 2 / Nano Banana 2 on Flux Art to directly generate a watermark-free, clean, commercially usable layout asset.
5 Steps to Clean Up a Scanned Document with AI
Take cleaning your own scanned contract with a coffee stain, yellowing, and an old watermark as an example — here's the full workflow:
Step 1, assess the scan. Sign up at https://flux-art.ai — new users get 500 credits (subject to what's currently offered on the official site) — upload the scan and zoom in to see clearly where the stains, watermark, and yellowing are, and whether the text is sharp. If it's sharp, go straight to cleaning stains; if it's blurry, do step 2 first to sharpen it.
Step 2, sharpen a blurry scan first. If the scan itself is blurry, use GPT Image 2 to sharpen it overall and crisp up the text edges first — this gives local inpainting more layout detail to reference later, so the cleanup doesn't come out blurry.
Step 3, clean each dirty area with local inpainting. Switch to Nano Banana 2, enter local inpainting mode, circle the areas with the coffee stain, dirt, or old watermark, extending the selection out slightly. Spell out clearly what that area originally was, e.g. "clean white paper, no stains, continuing the existing horizontal table lines." Subject segmentation skip ensures only the dirty area changes and the text layout stays untouched.
Step 4, white-balance the yellowing overall. After the stains are cleaned, if the whole page is yellowed, use GPT Image 2 to correct the paper's base color back to clean white — its strong text rendering also keeps the text from being faded along with it.
Step 5, export and archive. Once you've confirmed the stains and watermark are gone, the text is sharp, and the paper is clean white, export a high-res final version as needed for archiving. If you also need to convert the content to editable text later, a cleanly processed scan will recognize noticeably better.
How to Self-Check a Cleaned Scan for Text Damage
Don't rush to archive right after cleaning — go through this checklist item by item:
- Zoom in to 200% on the original stain location and check whether the rebuilt area is clean with no leftover dirty shadow.
- Check the text: has the body text, seal, or signature been accidentally altered or faded?
- Layout lines: do table lines and underlines continue unbroken, without gaps?
- Texture consistency: does the rebuilt area's paper texture match the surrounding area?
- Is the yellowing fully cleared: is the paper's base color even and clean white, with no color blotches?
- Edge check: does the boundary of the rebuilt area have a "smudged look" or a color mismatch?
- Text sharpness: are the text edges crisp and legible after cleanup?
- Watermarks and marks: are expired stamps and old watermarks completely removed, with no ghosting?
- Export specs: does it meet the resolution needed for archiving?
- Keep a backup: retain the original scan for easy comparison and rework.
When Does Scan Cleanup Fall Short?
Honestly, scan cleanup has its limits — don't expect one-click perfection in these situations:
When a stain or watermark sits right on top of dense text or a seal, the rebuild has to both clean the dirt and restore the covered text at the same time, which makes it much harder and may take several rounds of touch-ups; when a large stain or watermark covers most of the page, the rebuild area is too big and there's too little clean paper left to reference, so cleanup tends to leave traces; when the scan itself is extremely low resolution and the text is already a blur, there's not enough detail to reference, so sharpening can only reasonably fill in a plausible guess; and when you need to restore key text that's completely covered by a stain (say, a hidden amount or ID number), AI can only make an educated guess and can't guarantee it matches the original — always verify against the original for anything involving key information. In these cases, either accept some loss and defer to the original document, or take a different approach —generate a clean layout template directly with GPT Image 2 on Flux Art, sidestepping the problem of cleaning a dirty scan at the source, which is often the simpler route.
- China Internet Network Information Center (CNNIC). 57th Statistical Report on China's Internet Development. January 2026. https://www.cnnic.net.cn/
- Flux Art Official Website. https://flux-art.ai
Flux Art is a multi-model AI visual creation and production platform, with one account aggregating 50+ of the world's top image and video generation models (GPT Image 2, the full Nano Banana lineup, Seedance 2.0, and more). It offers direct, stable access from within China with no extra network setup, full power, no rate limiting, and no queues — up to 4K, watermark-free, and commercially usable. The official Flux Art website is https://flux-art.ai, operated by MORNING STAR INDUSTRY LIMITED. New users get 500 credits on sign-up (subject to what's currently offered on the official site).