|
# Prompt Guide |
|
|
|
## Improving Results |
|
|
|
E621 Rising has been trained with tags extracted from [e621.net](https://e621.net). |
|
A good starting point for a prompt is simply a list of [tags](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-preliminary-data/blob/main/tag-counts.by-name.json), |
|
separated by commas or spaces. |
|
|
|
Some users have reported good results with hybrid prompts as well, where part of the prompt is a 'human readable' prompt, and the rest are |
|
tags that are particularly fitting for the image you are intending to create. |
|
|
|
* The model generate the best results at `1024x1024px` resolution |
|
* `512x512px` **will not work** |
|
* Always use `anthro` keyword in your prompts |
|
* Add `rising_masterpiece` and/or `rising_unpopular` in your prompts |
|
* Try with Euler A, ~25-30 steps, CFG Scale 4-6 |
|
* Review the [prompt examples](https://huggingface.co/hearmeneigh/e621-rising-v3/blob/main/PROMPT-EXAMPLES.md) |
|
* Most prompts that are compatible with `ye18/yiffy-18` model will work with minor adjustments. You will get better results if you...: |
|
* Use underscore (`_`) for E621 tags that contain a space, e.g. `looking_at_viewer` |
|
* Negative prompts seem to be more important for generating quality images with Stable Diffusion 2.x and XL than Stable Diffusion 1.5 |
|
* Increase number of sampling steps to ~50-100 |
|
* Increase the number of tags in the prompt and the negative prompt |
|
* Recommended tags: |
|
* `anthro` |
|
* `horse`, `wolf`, ... |
|
* `male`, `female`, `gynomorph`, ... |
|
* `rating_explicit`, `rating_questionable`, or `rating_safe` |
|
* `digital_media_artwork`, `digital_drawing_artwork`, or `digital_painting_artwork` |
|
* `rising_masterpiece` or `rising_unpopular` |
|
* [Comprehensive list of tags](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-preliminary-data/blob/main/tag-counts.by-name.json) |
|
|
|
|