Article

How to Remove Image Watermarks with GPT Image 2 or Nano Banana

You have a photo with a corner logo, a diagonal text stamp, or a semi-transparent overlay blocking the scene — and you want it gone without smearing the background into a blur. GPT Image 2 and Nano Banana (Google's Gemini image models) can do this through semantic image editing: the model reads the surrounding pixels and reconstructs what should sit underneath the watermark.

What counts as a visible watermark

Corner logos — semi-transparent brand marks in a fixed position
Text overlays — "SAMPLE", "PREVIEW", diagonal copyright strings
Tiled or repeated stamps — preview grids across the whole frame
Platform export marks — social-app or stock-site badges baked into the pixels

These are composited on top of the image. AI editing models treat them as foreign objects and inpaint the region with contextually plausible background — skin texture, sky gradient, fabric weave, wall paint, and so on.

How AI watermark removal works

Modern image editors like GPT Image 2 and Nano Banana use semantic inpainting, not simple blur or clone-stamp. When you attach the watermarked photo as a reference and describe the edit, the model:

Parses the full scene — lighting, perspective, materials, and composition
Identifies the watermark region from your prompt (position, shape, opacity)
Synthesizes new pixels that match the surrounding context
Returns a cleaned image while keeping the rest of the frame as stable as possible

This is reconstruction, not pixel-perfect restoration. For marketing visuals, social posts, and internal mockups the results are usually excellent. For forensic or evidence-grade images, inspect the repaired area carefully — diffusion-style models can occasionally invent fine detail.

GPT Image 2 vs Nano Banana — which to pick

Scenario	Recommended model	Why
One-click cleanup from an existing image node	GPT Image 2	HiArt's AI edit action routes through GPT Image 2 by default
Complex textures — hair, fabric, architecture, product labels	GPT Image 2	Stronger instruction following and up to 4K output for fine detail
Quick iteration — try 3–5 prompt variants fast	Nano Banana 2	Flash-speed generation at lower credit cost
Large tiled watermarks or multi-region cleanup	Nano Banana 2 first, then GPT Image 2	Draft with Nano Banana 2; polish the best result with GPT Image 2
Final client deliverable after cleanup	GPT Image 2 or Nano Banana Pro	Higher fidelity and reference adherence on the last pass

Both models support reference-guided editing on HiArt: upload the watermarked image, describe what to remove, and generate. See the GPT Image 2 model page and Nano Banana 2 model page for specs and credit estimates.

Method A — AI edit with GPT Image 2 (fastest)

This is the quickest path when you already have an image on the HiArt canvas.

Upload or paste the watermarked image onto the canvas as an Image node.
Select the node and click AI edit in the toolbar.
Describe the watermark location and what to restore. Be specific about position and background type.
Generate. HiArt creates a new image-generator node with your photo as the reference and GPT Image 2 as the model.
Compare side by side. If a texture looks soft, re-run with a more detailed prompt or bump resolution to 2K / 4K.

Method B — Manual workflow with Nano Banana

Use this when you want to iterate quickly or prefer Google's image model for a particular scene type.

Add an image generator node and select Nano Banana 2 (or Nano Banana Pro for finals).
Attach the watermarked photo as the reference image.
Write a cleanup prompt (templates below). Match the source aspect ratio.
Generate at 1K or 2K for drafts; upscale or switch to Pro for the selected result.
Optional: pass the cleaned output back through GPT Image 2 AI edit for a final polish.

Prompt templates by watermark type

Name the watermark's position, shape, and the background you expect underneath. Vague prompts like "remove watermark" leave too much room for the model to guess.

Corner logo (bottom-right)

Remove the semi-transparent logo watermark in the bottom-right corner.
Rebuild the underlying background — match the surrounding sky gradient,
wall texture, and lighting exactly. Keep every other part of the image
unchanged: subject, colors, composition, and sharpness.

Diagonal text stamp

Erase the diagonal "SAMPLE" text watermark across the center of the image.
Restore the original photo content underneath — preserve skin tones, fabric
weave, and natural shadows. Do not blur or soften the subject. No new
text or logos.

Full-width banner overlay

Remove the horizontal text banner at the bottom of the image (approximately
the lower 15% of the frame). Reconstruct the ground, grass, or pavement
that should continue below the subject. Match perspective and color grade
of the rest of the photo.

Tiled preview grid

Remove all repeated preview watermarks across the entire image. Rebuild
each obstructed region with consistent texture — if the background is a
white studio sweep, restore a clean seamless white; if outdoor foliage,
restore natural leaf and branch detail. Keep the main subject identical.

Small logo on a product shot

Remove the small watermark logo on the product surface in the lower-left.
Restore the matte plastic / metal finish with correct specular highlights
and label edges. Do not alter product shape, branding on the product itself,
or studio lighting.

Tips for better results

Start at 2K — enough detail for the model to read texture; upgrade to 4K only after the composition is clean.
One watermark region per pass — if the image has both a corner logo and a center stamp, fix them in two generations.
Describe what should appear, not only what to delete — "restore blue sky with soft clouds" beats "remove logo".
Keep the source resolution high — heavily compressed JPEGs give the model less texture to work with.
Branch nodes on HiArt — keep the original watermarked image and A/B test prompt variants side by side.
Use PNG when possible — lossy compression artifacts around watermark edges make inpainting harder.

Common failure modes

Problem	Likely cause	Fix
Blurry patch where the watermark was	Prompt too vague or mask region too large	Add background description; split into smaller regions
Invented texture (wrong brick pattern, fake text)	Diffusion hallucination on complex areas	Re-run with "match exact surrounding texture"; try GPT Image 2 High quality
Subject face or product label changed	Watermark overlapped important detail	Narrow the prompt to watermark only; use two-pass cleanup
Color shift across the whole image	Model reinterpreted global color grade	Add "preserve original color grading and exposure exactly"
Watermark still faintly visible	Semi-transparent edges not fully described	Specify "including semi-transparent edges and ghosting"

Usage and legal note

Watermark removal is appropriate when you have the right to use the image — your own exports, properly licensed stock where clean masters are unavailable, or internal drafts with expired marks. Removing watermarks from third-party copyrighted work without permission may violate law or platform terms. When in doubt, license the clean original instead.

Ready to try it? Upload a watermarked image on HiArt canvas, hit AI edit, and start with the corner-logo template above. For deeper prompt craft on GPT Image 2, read our GPT Image 2 prompt guide. For high-volume iteration before a final polish, pair Nano Banana 2 drafts with a GPT Image 2 finish pass.