You look at the bottom corner of your beautifully generated image and see a smudge that looks like a watermark — usually unreadable text, a Shutterstock-style logo silhouette, or what looks like a signature in a corner. Sometimes it’s a clearly visible “shutterstock” or “getty” word; more often it’s a vague watermark-shape ghost.
This is training-data residue. The model was trained on millions of stock photos that had watermarks, and it learned that “professional photo” sometimes comes with a watermark-shaped artifact. It’s reproducing the shape even though there’s no real watermark behind it.
Common causes
Ordered by hit rate, highest first.
1. No negative-prompt watermark blocker
Without watermark, signature, logo, text overlay in the negative prompt, the model freely generates them when it thinks “professional photo” warrants it.
How to spot it: check your negative prompt. Missing these → that’s the cause.
2. Style words that evoke stock photo training data
"stock photo style, professional photography, commercial photography,
high quality photograph, editorial photography, cinematic"These words are over-represented in watermarked training data. The model associates them with the watermark itself.
How to spot it: prompt has stock photo, professional photography, editorial, commercial. Replace with neutral terms.
3. Checkpoint has heavy stock training
Some checkpoints (older SDXL fine-tunes, certain “photoreal” merges, models trained on web-scraped data) show more watermark residue.
How to spot it: same prompt on a known-clean checkpoint (Juggernaut XL v9+, RealVisXL v4+, Flux dev base) produces no watermark → original model is the source.
4. Generated at a resolution where corners get messy
Some models produce more corner artifacts at certain resolutions (e.g., 768×768 on SDXL). The center is fine, corners get garbage.
How to spot it: artifact is always in the same corner; center of image is clean. Try a different resolution.
5. Using a brand/style prompt the model associates with watermarks
shot like a Getty Images photo, Reuters style, documentary press photography — all of these have brand association in training data, and the model may reproduce the brand watermark.
How to spot it: prompt names a watermarked-source brand or style. Drop the brand name.
Shortest path to fix
Step 1: Negative-prompt watermarks (SD-family)
watermark, signature, logo, text, text overlay,
stamp, copyright, branding, getty images, shutterstock,
adobe stock, alamy, watermark text, attribution text,
artist signature, frame border, photo frameThis single addition removes the watermark in 70-80% of cases.
Step 2: Drop watermark-evoking style words
Remove from prompt:
stock photo, stock photography, professional photography,
commercial photography, editorial photo, getty style,
shutterstock style, reuters style, AP photoReplace with:
photograph, 35mm film, natural photo, documentary,
candid moment, available lightStep 3: Crop the outer 5% and re-upscale
Watermarks usually hide in corners. A simple crop + upscale fixes most:
# Cropping in Photoshop / Pixelmator
1. Crop ~5% off each edge
2. Resize back up to original dimensions (use bicubic or AI upscale)
3. Done — watermark is gone
# Command line (ImageMagick)
magick input.png -gravity center -crop 90%x90%+0+0 -resize 1024x1024 output.png
# Or just compose final layout with intentional crop
"... --ar 4:5" # Midjourney — taller crop hides corner artifactsStep 4: Inpaint over the watermark
If crop isn’t acceptable:
# Mask the watermark region
- Photoshop / Affinity / Pixelmator: Generative Fill / Remove tool
- SDXL Forge: inpaint with prompt matching surroundings
- Photoroom / Cleanup.pictures: web tools, free for one-off useStep 5: Switch checkpoint
Try the same prompt on:
# Known cleaner
- Juggernaut XL v9+
- RealVisXL v4+
- Flux dev base
- Midjourney (better filtered)
# Known watermark-heavy (avoid for stock-style)
- Older SDXL base
- Generic "photoreal" merges
- Some web-scraped fine-tunesStep 6: Use a higher-quality preset / quality flag
# Midjourney
"--quality 1" (default --quality 0.5 sometimes shows more artifacts)
# SDXL
- More steps (30-40) + better sampler (DPM++ 2M Karras)
- Higher CFG (6-7) — model commits more to "no watermark" if you specified itPrevention
- Bake
no watermark, no signature, no text, no logointo your default negative-prompt template — always paste it - Avoid
stock photo/professional photographystyle words; usephotograph/35mm filminstead - Pick known-clean checkpoints (Juggernaut, RealVisXL, Flux dev) for commercial work
- Always check corners on every generated image before delivery