An InvokeAI node that supports text generation models (gpt-neo, llama, etc) to create Stable Diffusion prompts - by default, it works with both a small and large gpt-neo model I made, trained on InvokeAI/Compel-formatted prompts.
It runs the text generation model on your chosen InvokeAI GPU/hardware, and offers configurable sampling, as well as word replacement. It also provides the ability to set a limit on the amount of tokens generated, and to provide a 'split' prompt conjunction to Compel, using its and() feature. Currently, if enabled, it will do this when it suspects that the generated prompt is over 75-77 tokens and would be truncated otherwise.
Navigate to your nodes folder (in your InvokeAI 3.4+ root folder), and run
git clone https://github.com/gogurtenjoyer/nightmare-promptgen
Alternately, you can download a zip of this repo with the green button in the upper-right here, and unzip it in your nodes folder.
To update, just cd into the nightmare-promptgen folder in your InvokeAI nodes folder, and run git pull. Otherwise, you can just download nightmare.py from this page and replace the older one.
Add the node as you would any other, then connect its output to a Compel positive prompt node, or similar.
Note: the first time you use this node in your workflow, it'll download a text generation model.
The default model is the Nightmare Promptgen XL model cactusfriend/nightmare-promptgen-XL (1.5gb), but if you'd like to save memory and download time, feel free to try the original: cactusfriend/nightmare-invokeai-prompts (550mb). These models were specifically trained with InvokeAI's Compel prompt format in mind, but feel free to try other text generation models!
You can duplicate or rename nightmare.default.yaml to nightmare.yaml, and the node will load this file, instead of the default. It'll fall back to the default file if there's a problem with your customized nightmare.yaml, as well.
You can add Huggingface repo IDs (or local file paths) for textgen models to the Models section, and words and a list of their replacements to the Replace section.
The default word replacement behavior checks that the word is separated from others by whitespace or punctuation; for example:
- replacing 'ape' in "my ape friend" would work, but in "I love grapes", it wouldn't.
If you'd like to override the 'word check' behavior, you can surround your word in ^ signs. In this case, ^ape^ would replace the text in both examples above. There's an example of this in nightmare.default.yaml.
If you'd like to use the Split Prompt feature, please keep in mind that you should send its output directly to the Compel prompt node, so that no other 'joined in' text could mess with the and() formatting. It's fine to join text into Nightmare Promptgen's 'prompt' input, though. Also, keep in mind that quotation marks should be escaped like \" in this prompt starter (Nightmare Promptgen handles escaping its generated quotation marks but doesn't do this for its input). Update: you can use the Quote Escaper node (included) to do this automatically now.
Adjusting the min and max new tokens generated can alter generation time dramatically. Because of this, there's also a max time setting that'll let you set (in seconds) the amount of allowed generation time. It'll overrule the other settings - think of it as a 'failsafe'.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent