This is a quick CLI utility I threw together from other people's work (see credits below) to upsample Stable Diffusion outputs (although it should work nicely on other things too)
It optionally fixes faces, and uses deep learning to scale up all the images in a directory
It should be enough to put some input images into the inputs folder and then do
conda env create -f environment.yaml
conda activate smalldogbig
And then
python smalldogbig.py
You can specify the upsampler by doing --bg_upsampler {option}. Can be one of: None, swinir (default, x4), swinir_x2, realesrgan (x4), realesrgan_x2, realesrgan_anime, hat (x4), hat_x2, edt
- realesrgan_anime is fastest, gives nice results on anime and illustrations, removes fine detail
- realesrgan is middle
- swinir is much slower, gives the most details
- hat gives the best results on real photos, but is often worse than realesrgan or swinir on synthetic images
- edt is an alternative to hat - very similar (although not quite as good) on real photos, but much faster
- The x2 versions scale to x2 instead of x4 before processing. They're faster, but otherwise worse
You will need at least 6GB of VRAM most upscalers. You can try setting a smaller bg_tile for 4GB.
You can adjust face correction with --w {adjust} which is 0-1.
- 0 gives more correction at the expense of accuracy
- 1 tries to be more accurate at the expense of correction.
Use 0.7 - 0.9 for mostly good looking faces, or 0.2 for messed up faces.
All of the scalers will auto-download their weights, except EDT.
For EDT, download all the SRx4_EDTB_* weights from https://mycuhk-my.sharepoint.com/:f:/g/personal/1155137927_link_cuhk_edu_hk/Eikt_wPDrIFCpVpiU0zYNu0BwOhQIHgNWuH1FYZbxZhq_w?e=bVEVeW and put them in weights\EDT
Original, linear interpolated 4x for comparison
SwinIR + Strong Face Correction
This is primarily just CodeFormer altered to use SwinIR as the upscaler, and the CLI tool simplified for my use case.
This is derived from CoDeformer, which is CC-BY-NC-SA/4.0. This is therefore also CC-BY-NC-SA/4.0.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
Component licenses:
- Codeformer CC-BY-NC-SA/4.0
- BasicSR: Apache 2.0
- SwinIR: Apache 2.0
- HAT: MIT
- EDT: None listed
City.png example image is derived (a scaled-down crop) from https://unsplash.com/photos/wpU4veNGnHg