Ai-Generated | Danbooru

This tag applies to fully AI-generated images or fully AI-generated animation. That is, images or videos which underwent no edits or extremely minor retouches. For edited images and videos, see AI-assisted.

Fully AI-generated art is against the upload rules and should not be uploaded. AI-assisted artwork is allowed as long as it meets quality standards and it was made with sufficient human involvement.

Examples

Stable Diffusion (2022)

NovelAIv1 (2022)

Midjourney (2023)

nijijourney (2023)

DALL·E 3 (2023)

NovelAIv3 (2023)

NoobAI-XL (2024)

NovelAIv4 (2025)

GPT4o (2025)

PixVerse (2025)

Identification

Below are some common signs that something may be AI-generated. If you are unsure whether a post is AI-generated or not, ask in topic #22285.

Generic art style

Many models have a distinctive "samey" art style. You can browse ai-generated status:deleted to get a feel for this.

Note that having a generic-looking art style isn't necessarily evidence of AI. Some artists may naturally draw this way, or have had their style copied by AI.

AI resolutions

Certain image resolutions are commonly used by AI models, and less often used by humans. These include:

832×1216, 1024×1024, 1216×832 (used by Stable Diffusion XL and NovelAI's Opus-tier).
512×768, 768×1152, 768×512, 1152×768, 768x1152, 1248×1824, 1824×1248, 1536×1536.

In general, AI models cannot generate at a resolution where both height and width are not divisible by 8, although humans can upscale, crop, or edit the final output to make it any resolution.

EXIF metadata

If the file contains metadata such as AI prompts or model names, it's a sure sign it's AI. On the upload page, you can click the filesize beneath the image to see the EXIF metadata. Danbooru will automatically show a warning and tag AI-generated after upload if the metadata indicates it is AI-generated.

Note that it's possible for an artist to draw on top of an AI-generated base or use an AI-generated background, in which case the work may be considered AI-assisted instead of fully AI-generated.

Inhuman mistakes

AI models often make abnormal or unusual mistakes within the artwork, such as:

Anatomical nonsense, such as extra legs (post #5732513), fused fingers (post #5742855), fused legs (post #5735212), etc.
Strangely drawn objects/environments such as a book with its cover as another layer of pages (post #5738243), wooden floor conforming with the shape of a character's hair (post #5850417), poor line continuity/object permanence when an object is hidden or goes behind other objects (post #5779922), etc.

Note that bad hands or other artistic errors are not in themselves signs of AI-generation, as humans can make mistakes too. The errors made by AI tend to be completely illogical or physically impossible things that no human would draw.

Large volumes of art

A brand new artist pumping out improbably large volumes of high quality-looking art in a short period of time, faster than any human could produce, is a strong indicator of AI. This isn't automatically a sign of AI-generation though, as sometimes new artists may release a pre-prepared portfolio of artwork to establish their presence on a new platform.

Unnatural art style changes

An artist with an unnatural change in art style after early-October 2022 (when NovelAI released its first image generation model) may be a sign they switched to using AI. This can easily be seen with artists with a low quality art style suddenly shifting into the generic-looking art style typical of NovelAI.

Some other notable dates for this would be:

2023/11/15: NovelAI V3 release
2024/10/18: NoobAI-XL release
2025/02/28: NovelAI V4 release
2025/03/26: GPT4o native image generation release
2025/05/29: NovelAI V4.5 release

Bear in mind that humans can naturally change their style or improve over time too. This would be for unnatural changes in style.

Garbled text

Older models have trouble generating realistic text, often producing gibberish or nonsense instead. Newer models may be better at generating text.

A lack of any signatures, web addresses, or watermarks can also be a clue, as human artists often sign their work, while fully AI-generated works often lack signatures.

AI mentions

Mentions of using "AI", software such as "ComfyUI", or common models such as "Nijijourney", "Midjourney", "NovelAI", or "Stable Diffusion" (also "SD" or "SDXL") in the tags, description, or the artist's bio are a very likely sign of AI.

Be careful not to confuse prohibitions against AI training with admissions of AI use. Counterexamples:

AI学習は禁止 (AI training prohibited)
AI学習厳禁 (AI training strictly prohibited)
AI不使用 (Not using AI)

Hidden Color Jitter / Latent Noise

AI models often produce tiny, unnatural RGB jitter in smooth areas (skin, sky, solid backgrounds, etc.) that is invisible to the naked eye.

To reveal this, apply an extreme RGB curve (a "wave curve" with 3 to 4 full oscillations from black to white across the 0-255 spectrum, e.g., 0-255-0-255-0-255-0) in any software that supports custom curves. After applying, AI-generated images often break into irregular color blocks, noisy grain, or strange patterns. In contrast, raw hand-drawn images remain relatively smooth or show natural gradients.

Note that this method may fail on very low-quality, extremely watermarked, heavily compressed, or heavily overpainted / AI-assisted images. It is a visual aid for human judgement, not a definitive test.

SynthID

Google and OpenAI models embed digital watermarks into their outputs that are invisible to the human eye, but can be read by their respective SynthID systems to determine if the image was edited or generated with those models. Note that false negatives are possible if the image was sufficiently tampered with or created prior to the implementation of SynthID, and SynthID cannot detect images originating from locally run or other companies' models.