Update README.md
Browse files
README.md
CHANGED
|
@@ -17,3 +17,18 @@ configs:
|
|
| 17 |
- split: train
|
| 18 |
path: data/train-*
|
| 19 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
- split: train
|
| 18 |
path: data/train-*
|
| 19 |
---
|
| 20 |
+
|
| 21 |
+
This dataset is basically 3500 "cherrypicked" images from my Civitai Gallery. <br>
|
| 22 |
+
From the oldest images to the newest. Using very short BLIP caption. On purpose lol yes. <br>
|
| 23 |
+
So from lamdba SDXL (...slop is hard word, let's say not great images xD...) as I take for 'variety' (to delete if too much sdxl plastic style biased for you until you reach good ones), to more decent SDXL refined image, and finally my current FLUX/KREA images for an interesting aesthetic style.<br>
|
| 24 |
+
Last SDXL upscaled images are actually the ones with better details when zooming. Was using A1111 API at the time to generate. <br>
|
| 25 |
+
The Flux/Krea images are now using Forge and involved many changes, resulting in a new compromise/style. I sacrified the 'realistic' aspect of the fine details when zooming, for a new style.<br>
|
| 26 |
+
So images fine details may qualified of 'sketchy' when zooming, but it's on purpose as the overall image when not zooming is perfectly the style I also wanted. <br>
|
| 27 |
+
To the final image, before posting, I also apply an adaptive color correction to push saturation, a kinda strong sharpening to unblur and reveal many details and a downsize to 70% to sharpen the image hen not zooming.<br>
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
I'm currently testing some trainings params. I'm posting the last """""decent""""" version here (totally a fail so far, but funny to play with!):<br>
|
| 31 |
+
https://civitai.com/models/2014817/wizzyfier<br>
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
|