The purpose
I’m trying out Pastel-Mix, a recommended Stable Diffusion model for 2D image creation.
I couldn’t get good images with the initial settings, so I’m leaving a note about it.
The image below was created with the settings from this article.
Download model and VAE
Download model
Please download the model from the following page:
For details on how to download, please refer to the article below:
It looks like the Hugging Face link has been removed.
Please download it from Civitai.
Download VAE
Download kl-f8-anime.ckpt from the following page.
Place the downloaded kl-f8-anime.ckpt file into the webui\models\VAE folder, located under your Stable Diffusion WebUI’s execution directory.
I was a bit worried that the “f8” in the filename might indicate it wouldn’t work with DirectML, but it ran without any issues.
Setting
Basic setting
Set the following:
| Checkpoint | pastelMixStylizedAnime_pastelMixPrunedFP16.safetensors |
| VAE / Text Encoder | kl-f8-anime.ckpt |
| Sampling method | DPM++ 2M |
| Sampler | Schedule type |
| Sampling Steps | 20 |
| CFG Scale | 7 |
Please make the settings within the red box. (The image below is a screenshot from Forge, so the actual screen layout may differ.)
Hires. Fix (Upscaling) Settings
To perform upscaling, use the following settings. (It’s generally recommended to use Hires. Fix for creating larger images rather than simply increasing the Width/Height.)
| Hires. Fix | Check |
| Upscaler | Latent |
| Hires Steps | 20 |
| Denoising Strength | 0.6 |
| Upscale by | Zoom ratio |
Please adjust the settings within the red box. (The image below is a screenshot from Forge, so the actual screen layout might differ.)
Execute
To execute it, simply click the Generate button, just like with other models.
Result
I was able to create a great image as shown below.
The prompt used was “a girl,” and the Seed was 100. You should be able to reproduce the exact same image by using these settings.
Appendix
This is the output when not using the recommended settings.
VAE：vae-ft-mse-840000-ema-pruned.ckpt
The image turned out a bit dark.
VAE：NONE
The image had low contrast.
Sampling Steps：10
It surprisingly produced an output, but details like the hands aren’t properly drawn.
Sampling Steps：40
The image changes slightly, but neither is particularly better than the other.
Since the generation time also increases, I feel like there aren’t many advantages.
Sampling method： Euler a
An astonishingly different image was created.
It’s so different that it’s unclear which one is better.
Schedule type： simple
You can tell the difference by comparing them. It might just be a coincidence, but in this example, the recommended settings seemed to produce more three-dimensional results.
comment