Spaces:
Running
on
Zero
Running
on
Zero
File size: 1,967 Bytes
0d34381 ae18532 0d34381 ae18532 3e63709 ae18532 dec8492 ae18532 3e63709 ae18532 0d34381 ae18532 3e63709 ae18532 0d34381 dec8492 0d34381 ae18532 0d34381 ae18532 0d34381 ae18532 0d34381 ae18532 fe94951 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 |
## Usage
TL;DR: Enter a prompt or roll the `🎲` and press `Generate`.
### Prompting
Positive and negative prompts are embedded by [Compel](https://github.com/damian0815/compel). See [syntax features](https://github.com/damian0815/compel/blob/main/doc/syntax.md) to learn more.
#### Weighting
Use `+` or `-` to increase the weight of a token. The weight grows exponentially when chained. For example, `blue+` means 1.1x more attention is given to `blue`, while `blue++` means 1.1^2 more, and so on. The same applies to `-`.
Groups of tokens can be weighted together by wrapping in parentheses and multiplying by a float between 0 and 2. For example, `(masterpiece, best quality)1.2` will increase the weight of both `masterpiece` and `best quality` by 1.2x.
### Models
* [cyberdelia/CyberRealisticXL](https://huggingface.co/cyberdelia/CyberRealsticXL)
* [fluently/Fluently-XL-Final](https://huggingface.co/fluently/Fluently-XL-Final)
* [segmind/Segmind-Vega](https://huggingface.co/segmind/Segmind-Vega) (default)
* [SG161222/RealVisXL_V5.0](https://huggingface.co/SG161222/RealVisXL_V5.0)
* [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
### Scale
Rescale up to 4x using [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) with weights from [ai-forever](ai-forever/Real-ESRGAN). Necessary for high-resolution images.
### Advanced
#### DeepCache
[DeepCache](https://github.com/horseee/DeepCache) caches lower UNet layers and reuses them every _n_ steps. Trade quality for speed:
* `1`: no caching (default)
* `2`: more quality
* `3`: balanced
* `4`: more speed
#### Refiner
Use the [ensemble of expert denoisers](https://research.nvidia.com/labs/dir/eDiff-I/) technique, where the first 80% of timesteps are denoised by the base model and the remaining 80% by the [refiner](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0). Not available with image-to-image pipelines.
|