santa hat
deerdeer nosedeer glow
Sign In

always trigger a word without writing it?

is it possible to inject/merge a prompt and negative prompt in a LoRa or Checkpoint so it is triggered by default so no need for the user to enter it every time?

like just inject it without training or something like that

2 Answers

If something shows up consistently in the image but isn't prompted it will get baked in.

Best example is how you should exclude hairstyle and color prompts when making a character LoRA (unless there are multiple variants). That way the model will not associate the attribute with any prompt.

You can also merge prompts if they show up often enough together by replacing them with one specific prompt - people often do that for clothing - but it is a bit fiddly as you need the right number/repetitions of each subset of images and the model loses a lot of flexibility.

Not sure how you would include negative prompts. Just make sure that the dataset doesn't have any (repeated) wrong tags. Specific negative prompts shouldn't be necessary then.

No, you can't inject a prompt into a model.

If what you're looking for is to always have the same positive and/or negative prompt for a given model, the quick solution I'd use, assuming you're using auto1111 webui, is to create Styles. It basically stores in a Style text for both positive and negative, and you can use multiple Styles simultaneously. That's how I use it.

Check Auto1111 github for more info:

Your answer