(artwork in the style of [name|name]:1.5),
(an image:0.5),
(:1.0)
copy this prompt
🔍 search
📓 features
📸 prompts
💫 about
📤 export
ESC
toggle all tags
see some random tags
copy visible artist names
Tap an artist's...
name to copy them to clipboard
tags to copy them to clipboard
⭐️ to toggle them as favorited
🌐 Bing search for the artist
🎨🧑🏞️ rotate between the 3 prompts
model switch between vanilla SDXL (A) and another model (B)
B: Model
Allows you to select a second XL model (B) to compare with the vanilla XL model from Stability (A).
Switch between A and B by tapping the matching button on an artist.
These models were selected by me based on what I prefer. I have no relationship with these model makers.
DynaVision:
Makes cute faces, removes artifacts, saturates colors, while preserving artist style.
Makes cleaner and smoother images at the expense of grungier styles.
For landscapes, it reduces artist style strongly but improves composition.
Improves anatomy and leans towards NSFW. Proportions are somewhat cartoonish.
To remove the "big eyes" effect, use "anime" in the negative prompt.
Crystal Clear:
Adds detail and removes artifacts while while strongly preserving artist style.
Preserves styles even for landscapes and even for messy/grungy artists.
Makes faces more aesthetic, improves anatomy, and leans towards realism.
Permissive filter
checked: artists matching any checked tags
unchecked: artists matching all checked tags
useful for filtering your favorites
Hide low-use tags
checked: hides tags that match less than 4 artists, ~50% of all tags
you can access hidden tags via search, or if they're marked as important
Hide unknown to SDXL
I've hand verified that SDXL doesn't know these artists' styles
If you prompt with their names, the result will be unlike their actual style and generic, and you can achieve similar results by prompting "a painting", etc.
They're included in the database for the record of what SDXL doesn't know
Don't blur risqué
Risqué means any image that isn't G-rated and 100% SFW by extremely conservative standards. But the most risqué this site gets is images with boobies.
Important tags
Tap edit under the important category to add or remove tags to that category. This lets you save your "favorite" tags for easier access.
Build a prompt
See 📸 prompts
But wait, there's more...
Quickly toggle any tag by typing the / key, then typing the first few letters of the tag
Need inspo? Try the "see some random tags" button from the search tab. Does what it says on the tin
"Copy visible artist names" from the search tab lets you export all artists that match you filters as plain text
The export tab lets you copy data from one device to another
Why does the 🎨 prompt sometimes match the 🧑 or 🏞️ prompt?
The prompt used for the 🎨 images is the intentionally generic, "an image". If that results in a portrait, that's useful to know. It indicates that, for that artist, SDXL is more strongly trained to output portraits. Anything in your prompt that's not related to a person will harder to prompt for in that artist's style.
✍️ Offline tag-editing mode
This feature is only for editing tags offline. If you are viewing this from Hugging Face, any tag edits won't be applied, but they'll be saved. If you've downloaded the repository and are viewing your local file, tag edits are applied. In both cases, your edits will be available for export 📥 as text. If you want to suggest tag improvements for everyone's benefit, edit tags, copy the text from 📥, and post it as a comment to Hugging Face. I will incorporate accurate edits ASAP. Thank you for contributing!
Building prompts from artist styles
Build-a-prompt uses auto1111 syntax, and it creates a 3-section prompt:
i.e. ✖1, ✖2, ✖3, gives more weight to a style by repeating the artist's name. This works better than using prompt weighting
Mixing method:
Add: puts all the artists' names in the prompt separated by a comma. Usually the best choice. Try 'add' and 'loop' and see which works better.
Loop: uses a1111 syntax for alternating the prompt at each step. This alternates which artist's name appears in the prompt. If an artist's name is repeated, then it's repeat for multiple steps.
Swap: uses a1111 syntax to swap the prompt after a fraction of the total steps. This swaps the artist's name after 50% of steps by default. You can change this fraction. For example, [A:B:0.1] switches to B after 10% of all steps.
Move left and right:
tokens closest to the start of the prompt always have more influence. Therefore, moving an artist's name to the left adds weight to that style.
Why this prompting method?
The reason why build-a-prompt uses 3-sections is that it's very easy for common words in a prompt to erase an artist's style. For example, for some artists that never made any portraits, using the word "portrait" in your prompt can erase their style entirely.
To overcome this, we add weight to the artist stuff and to remove weight from the rest of the prompt. Built a prompt weights this at 1.5 and 1.0 respectively. This gives slightly more weight to the artist stuff. As you edit the rest of the promt, if the style disappears, adjust both numbers, e.g. try 2.0 and 0.5.
The middle section, "(an image:0.5)" doesn't change the output at all. However, it's sometimes needed when the weight of the rest of the prompt weight is less than 1.
An artist's style includes much more than their medium
Using an artists name influences the output in several ways beyond the medium, e.g. "painting style". It includes typical subject matter, colors palette, clothing and hairstyles of the artist's time, composition and layout, even specific faces of people, and other aesthetic choices.
It can be hard to get just the aesthetic style and not the other stuff, e.g. a renaissance painting style with modern clothing.
Sometimes you can strengthen just the artist's style by putting their medium in the prompt like this:
(watercolor in the style of artist full name:1.5), (an image:0.5), (content stuff:1.0)
This helps especially when the artist is famous for multiple mediums. But sometimes it reduces the artist's specific style and more matches the generic style of the medium.
Also try adding the opposite medium/style to the negative prompt. For example, if the artist you want is a photographer, try "painting, illustration" in the negative prompt.
Making non-photographic mediums look more photographic
It's worth trying "photograph by illustration artist" along with "illustration" in the negative prompt. But that often won't work.
It's also worth trying "[by artist A:style of photograph:0.5]". This will start with the artist style and switch to photograph style after 50% of the steps. Adjust the decimal as needed.
I've had best results by sending the output to img2img, then replacing the artist's name in the prompt with "photograph", and setting the de-noise very low. Then feed the result back to img2img, and slowly nudge it towards photographic. That will at least maintain the composition.
If Reference Controlnet for SDXL has been released, that should work well.
🎨 prompt: (artwork in the style ofartist full name:1.5), (an image:1.5)
negative:
edges, borders
🧑 prompt: (artwork in the style ofartist full name:1.5), (an image:0.5), (portrait of a person, inside a place:1.0)
negative:
edges, borders, outside, about, the artist, themselves
🏞️ prompt: (artwork in the style ofartist full name:1.5), (an image:0.5), (landscape, outdoor natural scenery, water, bridge:1.0)
negative:
edges, borders, inside, people, person
exceptions:
For most photographers, "artwork in the style of name" was replaced with "by name". Rarely, seed 48 was used. For Dynavision I used negative prompt keywords to reduce NSFW.
Disclaimers
The tags are mostly my manual input, with a mix of other sources including AI. They may be incorrect!
SDXL is only a crude imitation of these artists. Check out these artists' actual artwork for more inspiration!
This database has more straight white male artists than in the actual population. That's because they've been favored by art-collectors, past and present. Please suggest non white male artists that I should add.
My code doesn't use cookies, and it sends nothing to any server. But it's hosted on Hugging Face, and I can't control if they cookie you.
This tool is open source and Creatives Commons license. Download for free, use offline, modify as you wish.
I don't get, nor do I want, any compensation for this. Getting thanked feels great though!
I'm not affiliated with Stability AI or any model makers.
This tools is only for fun. Use at your own risk. You're responsible for your own actions. Void where prohibited.