Watermark appearing despite negative prompt tag
Hey,
Running into an issue with Anima preview — some generations are coming out with a visible watermark even when watermark is explicitly placed in the negative prompt.
Tested across multiple seeds and prompt variations. The tag is there, negatives are weighted, but certain artist styles / LoRA combos seem to bake the watermark in at model level regardless.
Thanks
pics or it didn't happen...
But seriously, how do you expect anyone to confirm your claims without posting the image+workflow/prompt? I have seen watermarks with trash-tier Lora where the trainer couldn't be bothered to sanitize the dataset, and I'm not doubting that you have the issue, but it must be very specific to either an artist or franchise that has notoriously prominent watermarks in the dataset, but without a way to replicate your results noone can help you.
pics or it didn't happen...
But seriously, how do you expect anyone to confirm your claims without posting the image+workflow/prompt? I have seen watermarks with trash-tier Lora where the trainer couldn't be bothered to sanitize the dataset, and I'm not doubting that you have the issue, but it must be very specific to either an artist or franchise that has notoriously prominent watermarks in the dataset, but without a way to replicate your results noone can help you.
positive prompt:
year 2025, newest, masterpiece, best quality, absurdres, highres,
@bartolomeobari,
1girl, yae_miko, genshin_impact,
fox ears, pink fox ears, multiple tails, nine tails,
long hair, light purple hair, pink hair, hair down,
purple eyes, red eyeshadow, eyeliner,
miko outfit, white dress, wide sleeves, sakura pattern,
gold hair ornaments, earrings,
sitting on shrine steps, one knee up, elbow resting on knee,
chin resting on hand, relaxed confident expression, slight smile,
looking at viewer,
japanese shrine, stone lanterns, cherry blossom trees,
late afternoon light, golden hour, petals falling,
warm golden pink light, long soft shadows,
three-quarter view, slightly low angle,
detailed beautiful eyes, fabric pattern detail, petal scatter,
warm gold pink theme, elegant, confident, shrine maiden
negativ prompt:
worst quality, bad quality, low quality, lowres,
bad anatomy, bad hands, extra fingers, missing fingers, extra limbs,
deformed, mutated, ugly, disfigured,
signature, watermark, username, artist name,
jpeg artifacts, scan artifacts,
multiple views, comic, 2koma, 4koma,
sketch, lineart, rough, unfinished,
oversaturated, washed out colors, waxy skin,
blur, blurry, blurry background, blurry foreground,
blurry face, blurry hands, motion blur, depth of field,
out of focus, soft focus, gaussian blur, bar_censor, censored, mosaic_censoring,
chromatic aberration, noise, grain, shiny_skin,
The gist of it is that AI will inevitably pick up commonly repeating patterns. If many images associated with an @artist tag contain that watermark, the model will associate the watermark with the tag. AI might know what a watermark means, but putting watermark in the negatives won't remove it from the image, because there is no strong separation of art style and watermark when it comes to @artist tag in the positive conditioning.
This is nothing specific to anima, SDXL booru models also suffered from the same issue.
You can try:
Adding similar tags to negatives like: patreon logo, artist logo, signature, twitter username, artist name etc.
Adding natural language instruction like "No watermark, no text in the image" to positives.
Adding natural language instruction like "There are visible watermark and text in the image." to negatives.
Increasing (strength:) of any of the above.
to see if they help.
Bartolomeobari is notorious for watermarks in datasets. Same with any illu/noobai model that is good at artstyle reproduction and not slopmerged to death.
That said, I think the deviantart watermarks are a little over the top... This should definetely be fixed.
The gist of it is that AI will inevitably pick up commonly repeating patterns. If many images associated with an @artist tag contain that watermark, the model will associate the watermark with the tag. AI might know what a watermark means, but putting watermark in the negatives won't remove it from the image, because there is no strong separation of art style and watermark when it comes to @artist tag in the positive conditioning.
This is nothing specific to anima, SDXL booru models also suffered from the same issue.
You can try:
Adding similar tags to negatives like: patreon logo, artist logo, signature, twitter username, artist name etc.
Adding natural language instruction like "No watermark, no text in the image" to positives.
Adding natural language instruction like "There are visible watermark and text in the image." to negatives.
Increasing (strength:) of any of the above.
to see if they help.
Added everything you wrote onixxexxd5555LOAF watermarks present
As I said, it was a suggestion. I do not know any 100% reliable solution to this problem, or if such exists. You could maybe also get rid of the bird at the bottom with twitter logo tag.
Run 10 seeds, such as 1 to 10, with your initial prompt. Then run the same seeds again with the modified prompt. Hopefully you should see a decrease in the frequency of how often watermarks appear, even if it doesn't reliably make them go away.
I think when referencing an artist not in the training data with "@artist name" to indicate style, the model sometimes treats it as literal text and generates it as a watermark.
I think when referencing an artist not in the training data with "@artist name" to indicate style, the model sometimes treats it as literal text and generates it as a watermark.
No, this is literally Bartolomeobari's signature on every image he posts. All models trained on booru scraps have these watermarks for Bartolomeobari, even if you try to remove them with negs.