90 detailed dynamic prompts with fantasy, sci-fi, video game, mythological and historical themes. Each prompt includes setting, background, pose and (mostly) gender-agnostic character details. Now with booru tags!
For ComfyUI, please use https://github.com/adieyal/comfyui-dynamicprompts, and https://github.com/GrvBdgr/comfyui-negativewildcardsprocessor.
The wildcards contain negative wildcards (text surrounded by <! .. !>). If you do not wish to use these, simply remove the <! .. !> text in the wildcards files with your favorite text editor. Use the following regex:
<!.*!>As of v3.0 the wildcards no longer require Loras or embeddings, for compatibility with all models.
Older versions of the wildcards are not recommended for use.
Usage:
Add a description of the character you'd like in the generated image (feel free to use character LoRas), plus any style/quality keywords necessary for the model you're using, then add __ffpc-booru__.
Example:
best quality, masterpiece, 1man, russian (male:1.1), solo, aquamarine eyes, long brown beard, __ffpc-booru__
Some example characters are provided in the character.txt in the downloaded archive, so if you just want to run a few tests, you can use the following prompt:
best quality, masterpiece, __character__ __full-prompt-fantasy__
More diverse character wildcards can be found here: https://civarchive.com/models/118508?modelVersionId=128565
Negative prompt should be whatever the model necessitates, or your preferred negative prompt. Some models have race or gender bias due to the material that was used for training, you can usually use negative prompts to compensate. eg. if your generations come out with female characters despite describing the character as male, add "woman, girl" to the negative prompt, use weights if necessary.
Additional usage notes:
Other than the above, feel free to experiment with additions to the prompt, though most of the wildcards are rather long, so if you want to have a significant impact, weights may be necessary. The prompts rarely use weights above 1.1, so if your additions seem to have no effect, try weights of 1.15 or more.
Description
Lite (Lora-less) version of v2.2. Please check description for details.
FAQ
Comments (15)
Do you have some favorite models that have a tendency to work well?
My response to an earlier, similar question:
My recommendations for non-realistic:
https://civitai.com/models/112857/hellocomic
https://civitai.com/models/54073/arthemy-comics
https://civitai.com/models/70692/lemontea-mix-painterly-25d
Realistic is a bit harder, as most photorealistic models aren't exactly trained with the themes of these wildcards in mind. The following produce decent results, and are more on the realistic side of the scale:
https://civitai.com/models/123777?modelVersionId=147577
https://civitai.com/models/84728?modelVersionId=90072
https://civitai.com/models/52298?modelVersionId=90072
Other than these, I'd recommend giving https://civitai.com/models/158837/restlessanimance a try, it's an underrated model for how well it works (especially with these wildcards).
Also, check out the sample images, and the gallery for the wildcards (even past versions), and see which styles you like. Most images have resource info uploaded along with them, so you can easily get to the model from them,
I had problems using the script. The download never finished and it would try 5 more times and then abort and move on to the next one. When I checked the Lora folder, I realized that there were some files, but all were corrupted. I got around this by modifying TIMEOUT_READ to 100, which stopped the download from aborting and downloaded all of them.
If you get constant fails with the default TIMEOUT_READ value, it means that civitai takes more than 10 seconds to respond to your download requests. This can mean that the civitai servers are busy, or that there is something disturbing your network traffic (bit torrent clients can do that for exampe).
Glad you could get it to work with TIMEOUT_READ = 100, but this is not an issue with the script, rather a sign of something inhibiting network traffic on your side, or civitai's side.
@TxcTrtl Maybe it's for geographical reasons. I changed TIMEOUT_READ=100 and creating a new constant RETRY_DELAY=10 and changing the line time.sleep(TIMEOUT_READ) to time.sleep(RETRY_DELAY). Before running the script I deleted all corrupted Loras from the folder because if this is not done it will ignore them as if they were already downloaded. That was enough and solved my problem.
the custome timeout code in lora_downloader.py is redundant, and breaks the download while it is still ongoing, requests already monitors the timeout when using it:
requests.get(url, stream=True, headers=headers, timeout=(TIMEOUT_CONNECTION, TIMEOUT_READ)) as response:
remove the custom timeout code and the script will work.
good work, and thank you.
The requests library indeed handles the timeout by itself, but I found that on occasion, civitai generates some sort of response, that doesn't trigger the native timeout that the library recognizes. I didn't investigate deeper than this, and instead elected for a workaround -> I added the custom timeout, to catch such errors and prevent the script from freezing. If the download is particularly slow, the custom timeout can break it while it is in progress, that much is true. This can be circumvented by increasing the TIMEOUT_READ from 10 to something higher. I'll modify it in the next version.
I wish I saw this before manually downloading every single lora
The fpf_download_links.txt that comes with v2.2 causes the newest loras to be saved without the .safetensors extension, making the loras not show up in the Lora tab. Adding .safetensors to the end of each name starting from Type3Tech-21 will fix this.
Thanks for pointing that out! Will be corrected in the next version.
Is possible to use with sdxl?
The LoRAs used in the wildcards are for SD1.5, so they wouldn't work with SDXL. The Lite versions don't include any LoRAs, so you should have no problem using them with SDXL.
About half the Lora links are outdated and do not work...
Which links did you find that are outdated?
I just ran a test of the downloader script that is included with v2.2, and all LoRAs were downloaded without issue.
@TxcTrtl many failed the first time but it looks like they eventually downloaded, maybe the server was over loaded. Sorry. You do awesome work, thanks!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.