The fastest fix is a strong text negative prompt. Add “text, watermark, signature, artist name, logo, username, caption, subtitle, letters, words” and weight it up. That clears most hallucinated text and phantom signatures in a single re-roll. Inpaint any stubborn mark that survives at denoise 0.5 to 0.7.
The image is clean except for the garbage the model invented: a fake artist signature in the corner, a phantom watermark across the frame, a subtitle bar of garbled letters, or a nonsense logo stamped on the wall. Your output is hallucinating text that was never in your prompt, and it wrecks an otherwise usable render. This is one of the most fixable AI defects, because it comes from a small set of clear causes.
This content is for adults 18+ and every character described is a fictional adult. This is about your own output inventing text, not about tools that stamp watermarks on you, which is a completely separate topic.
The important mental model is that the model is not trying to sign your art or protect it. It is reproducing a pattern it learned: pictures often have a mark in a corner, so it adds a mark-shaped smudge. Once you see it as a learned pattern rather than a deliberate stamp, the fix is obvious, since you can steer the model away from that pattern with the negative prompt and repair the rare survivor by hand.
Why AI images hallucinate text and watermarks
The fake text comes from a few clear sources, and knowing which one you are hitting decides the fix.
Watermarked and signed training data. Huge amounts of the art these models learned from carried signatures, watermarks, usernames, and site logos, especially booru-sourced and stock-sourced images. The model learned that images often have a mark in the corner, so it reproduces that pattern as generic scribble-text. This is the single biggest cause, and it is why the fake marks so often land in a corner or across the bottom, exactly where real watermarks live.
Checkpoints and LoRAs baked with signatures. Some finetunes were trained heavily on one artist’s signed work or on watermarked packs, so they compulsively add a signature-shaped smudge. If a specific model or LoRA always drops a mark in the same spot, it is baked into that model’s weights, and no amount of prompt cleverness fully removes it while that model is loaded.
Prompt words that imply text. Words like poster, magazine cover, album art, sign, logo, book, newspaper, or storefront tell the model the scene should contain writing. It obliges with garbled letters because it cannot spell. Your prompt may be inviting the text without you noticing, since these words feel like scene descriptions rather than instructions to render letters.
No anti-text negative. If your negative prompt says nothing about text, you are leaving the door open for every learned watermark pattern to appear. Absence of a text negative is itself a cause. A simple negative closes off the large majority of cases, which is why it is the first thing to add and the reason so many people never see the problem at all once they include it by default.
How to fix hallucinated text and watermarks
Work these easiest-first. Step one alone resolves most of it, so start there before touching models.
1. Add a strong text negative prompt
Put a full anti-text block in your negative: text, watermark, signature, artist name, logo, username, caption, subtitle, letters, words, writing, title, label. This directly steers the model away from the mark-shaped patterns it learned. It is the single most effective fix and costs you nothing but a re-roll. Keep this block in your default negative permanently, since the marks can appear on any prompt and the block does no harm when there was no text to begin with.
2. Strengthen the negative with weighting
If a faint mark still slips through, increase the weight on the text terms, for example (text:1.4), (watermark:1.4), (signature:1.3). Emphasis makes the model push harder against those patterns. If you are unsure how weighting syntax works, the prompt weighting guide explains the parentheses and numbers. Bump the weight gradually rather than jumping straight to extreme values, since overweighting the negative can start distorting textures where the model expected a mark. The full negative vocabulary lives in the negative prompts master list.
3. Avoid prompt words that summon text
Scan your positive prompt for words that imply writing: poster, magazine, sign, logo, cover, newspaper, book, storefront, billboard. If they are not essential, cut them. Removing a single “magazine cover” tag can stop the model stamping a garbled masthead across the top of the frame. Many booru-style prompts inherit these words from copied tag lists, so it is worth reading the whole prompt with fresh eyes and deleting scene words you never actually needed.
4. Inpaint over the mark
For a mark that survives the negative, mask just that region and inpaint with a prompt matching the surrounding area (skin, plain wall, background) at denoise 0.5 to 0.7. This regenerates the marked patch as clean content. It is the reliable fix for a single stubborn signature in the corner of an otherwise perfect image, and it beats re-rolling the whole render when everything else about the image is exactly what you wanted.
5. Drop or lower the offending LoRA
If the same signature appears every time you use one LoRA, remove it or lower its weight to around 0.5 and test. A signature-prone LoRA is fighting your negative. Reducing its influence lets the anti-text negative win. Test the LoRA in isolation first, with a clean prompt, so you can confirm it is the source before you start blaming your wording. Browse alternatives in the NSFW LoRA roundup if one is chronically dirty.
6. Switch to a checkpoint that is not signature-prone
When a specific model always marks its output regardless of your negative, the signature is baked into the checkpoint. Move to a cleaner base model. Some checkpoints are trained on heavily watermarked packs and will fight you forever, while others render clean out of the box. The checkpoint guide covers models that render clean without a compulsive corner mark.
7. Crop or clone it out
If the mark sits at the very edge, a simple crop removes it in seconds. For a mark in a flat area, a clone or heal pass in any editor paints it out. These are the pragmatic last steps for a render you do not want to regenerate at all, and for edge marks a crop is often faster and cleaner than any inpaint.
Cause to fix reference
| Root cause | What you see | The fix |
|---|---|---|
| Watermarked training data | Generic scribble in corner | Strong text negative prompt |
| Signature-baked LoRA | Same mark every use | Drop or lower LoRA weight |
| Text-implying prompt word | Garbled masthead or sign | Remove poster, magazine, sign |
| No anti-text negative | Random letters anywhere | Add full text negative block |
| Faint mark persists | Ghost watermark | Weight up the text negative |
| Edge or flat-area mark | Corner signature | Crop or clone it out |
Copy-paste settings
This is the full anti-text setup: a clean positive with no text-summoning words, and a heavily weighted anti-text negative. The inpaint block clears anything that survives the negative.
Sampler: DPM++ 2M Karras
Steps: 30
CFG: 5.5
Resolution: 832x1216 (SDXL)
Positive:
beautiful adult woman, detailed skin, soft lighting, sharp focus,
professional photography, clean plain background
Negative:
(text:1.4), (watermark:1.4), (signature:1.3), artist name, logo,
username, caption, subtitle, letters, words, writing, title, label,
stamp, copyright, lowres, blurry, child, minor, underage, loli, shota
Inpaint (for a surviving mark):
Mask only the mark. Denoise 0.55. Steps 25.
Prompt: matching background, skin, plain wall (no text)
Common mistakes
No text negative at all. The most common reason people fight hallucinated text is that their negative never mentioned it. A short text block is the first thing to add, not the last, and once it lives in your default negative the problem mostly disappears for good.
Fighting a baked LoRA with the negative. If a signature-prone LoRA keeps winning, no amount of negative weighting fully clears it. Lower or drop the LoRA instead of escalating the negative into absurd weights that distort the rest of the image. The negative cannot out-argue a pattern that is trained deep into the weights you loaded on purpose.
Leaving text-summoning words in the prompt. Keeping poster, magazine, or sign in your positive while adding a text negative sets the two against each other. Remove the invitation rather than only fighting the result, since it is far easier to not summon the text than to suppress it after the fact.
Inpainting at too high a denoise. Masking a corner signature and inpainting at denoise 0.95 can invent new content or a new mark. Keep it around 0.5 to 0.7 so the patch blends into the surrounding area and does not generate a fresh problem where the old one was.
Over-weighting the negative into distortion. Pushing (text:1.9) and similar can start warping textures and edges near where the model expected a mark. Weight up moderately, to around 1.3 to 1.5, and switch to inpaint if that is not enough. Distortion from an overloaded negative is harder to fix than the original mark.
Building a permanent anti-text default
The smartest long-term move is to stop treating hallucinated text as a per-image emergency and start preventing it by default. Once a solid anti-text block lives in your standard negative prompt, the problem quietly disappears from the vast majority of your renders and you only deal with the rare survivor. This is the single habit that separates people who constantly fight watermarks from people who almost never see one.
Build a base negative you paste into every generation. It should carry the full text family (text, watermark, signature, logo, username, caption, subtitle, letters, words, writing) alongside your usual quality and safety terms. Because these terms do nothing when there was no text to begin with, there is no downside to keeping them permanently. Treat them like a seatbelt: on every time, noticed only when they save you.
Pair that default with a short pre-flight habit of scanning your positive prompt for text-summoning words. Poster, magazine, sign, logo, cover, and storefront are the usual suspects, and they sneak in through copied tag lists more often than through deliberate choice. Deleting them before you generate is far cheaper than suppressing the garbled text they invite. A prompt that never asks for a scene full of writing rarely produces one.
Keep a mental note of which of your models and LoRAs are dirty. If you know from experience that a particular checkpoint stamps a corner every time, you can decide up front whether to avoid it for clean work or to plan on an inpaint pass. The same goes for a signature-prone LoRA: if it is worth using for its style, budget the extra cleanup step rather than being surprised by it. Knowing your tools’ habits lets you choose them deliberately instead of reacting to their output.
When a mark does slip through despite all of this, you already have the workflow: weight the negative up a notch, and if that is not enough, inpaint the region at a moderate denoise. Because your default already blocks most text, the survivors are few and usually small, which makes the occasional inpaint a quick touch-up rather than a recurring battle. Prevention at the prompt level plus a reliable repair for the exceptions is the whole system, and it keeps hallucinated text from ever being a real obstacle again.
Why the model cannot spell, and what that means for you
Understanding why the fake text is always garbled helps you set the right expectations and pick the right fix. Diffusion models learn the visual shape of letters and marks, not language, so they reproduce something that looks like writing without ever being able to spell a real word. That is why a hallucinated signature is a plausible scribble and a hallucinated caption is a row of letter-shaped nonsense: the model is painting the look of text, not text itself.
This matters because it tells you the mark is a texture the model adds, not a message it is trying to send, so you fight it the same way you fight any unwanted texture: with the negative prompt and with targeted repainting. It also tells you that trying to make the model produce correct text by prompting harder is a losing game on most Stable Diffusion checkpoints, since the capability simply is not there. If you actually need readable words in an image, add them in an editor afterward rather than hoping the model spells them.
The garbled nature of the marks is also why inpainting works so cleanly. Because the fake text is a shallow surface pattern rather than something woven into the structure of the scene, masking it and repainting with a matching background prompt removes it without disturbing the real content underneath. You are painting over a smudge, not surgically extracting a deeply embedded object, which is why a moderate denoise is enough and why the repair blends so easily.
Keep this model in mind and the whole problem shrinks to its true size: a learned decorative pattern that appears when nothing tells the model not to add it. Block it by default, avoid inviting it with text-summoning words, and repaint the occasional survivor. There is nothing mysterious about hallucinated text once you see it as the model imitating the look of a watermark it saw ten thousand times in training, and that framing points you straight at the fixes that actually work. Seen clearly, it is one of the easiest AI defects to control, because a single default negative prevents most of it, a quick word audit prevents the rest, and a moderate-denoise inpaint cleans up whatever rare mark still manages to slip through onto a finished render.
Verdict
Hallucinated text and phantom watermarks are a training-data artifact, and a strong, slightly weighted text negative clears the large majority in one re-roll. Remove any text-summoning words from your positive, drop signature-prone LoRAs, and inpaint the rare survivor at denoise 0.5 to 0.7. Switch checkpoints only if a model marks every single render regardless of your negative. Keep the text negative in your default prompt and you will rarely see the problem again.
Frequently asked questions
Why does my AI image add fake text or a signature I never asked for?
Because the training data was full of watermarked, signed, and captioned images, especially booru and stock sources. The model learned that pictures often carry a mark, so it reproduces that pattern as garbled scribble-text. Adding a strong text negative prompt steers it away from those learned mark patterns and clears most cases.
What negative prompt removes hallucinated text?
Use a full block: text, watermark, signature, artist name, logo, username, caption, subtitle, letters, words, writing, title, label. Weight the key terms up, for example (text:1.4), (watermark:1.4). This directly pushes the model away from the mark-shaped patterns it learned and resolves the large majority of hallucinated text in one re-roll.
How do I get rid of a signature that survives the negative prompt?
Inpaint it. Mask only the mark and regenerate that patch with a prompt matching the surroundings, like skin or plain wall, at denoise 0.5 to 0.7. This rebuilds the marked region as clean content while leaving the rest of your image untouched. For an edge mark, a simple crop is even faster.
Why does one LoRA always add a watermark?
Because that LoRA was trained on watermarked or signed images, so a signature-shaped smudge is baked into its weights. It fights your negative. Lower the LoRA weight to around 0.5 or drop it entirely and test. If the mark disappears, the LoRA was the source and the anti-text negative can now win.
Do certain prompt words cause text to appear?
Yes. Words like poster, magazine, cover, sign, logo, newspaper, book, and storefront tell the model the scene should contain writing, and it obliges with garbled letters because it cannot spell. Remove any text-implying words you do not actually need, since they invite the very text you are trying to avoid.
Can weighting the negative help clear phantom watermarks?
Yes. If a faint mark still slips through, increase the weight on the text terms, such as (text:1.4) or (signature:1.3), so the model pushes harder against those patterns. Keep it moderate around 1.3 to 1.5, since very high negative weights can start distorting textures near where the model expected a mark.
Is this the same as tools that stamp a watermark on my images?
No. This is about your own generation hallucinating fake text, signatures, and logos that were never in your prompt, caused by watermarked training data. It is a generation-side artifact you fix with a negative prompt and inpainting, not an external tool adding a stamp to a finished file.
Should I switch checkpoints to stop hallucinated text?
Only as a later step. First try a strong text negative and removing text-summoning words, which fixes most models. If one specific checkpoint marks every single render regardless of your negative, the signature is baked into that model and switching to a cleaner base checkpoint is the reliable answer.
