Update README.md
Browse files
README.md
CHANGED
@@ -1,15 +1,6 @@
|
|
1 |
-
Newer from-scratch merge here: https://huggingface.co/zatochu/JinkiesNoob
|
2 |
-
|
3 |
-
JinkiesNoob is a lightning-free merge from the ground up that aims to simply fix Noob 1.0 somewhat while also adding in rudimentary natural language prompting support. This was done by training multiple loras on multiple natural language datasets and baking in the CLIP/TE lora patches. Avoding UNET patches as that will influence style range far too much.
|
4 |
-
Use more typical sampling settings. 5-7.5 CFG + 20 steps and up.
|
5 |
-
Natural language understanding is pretty basic so don't expect to see super nuanced and flowery prompts working. But it does allow for some simple short prompts that conscisely describe what you want. Don't forget tags will still function.
|
6 |
-
|
7 |
-
Be aware that the non-Lightning merge will struggle more with clean detail. While ChromaXL Mix introduces that grain on flat color section, it still offers quite a boost in detail and convergence.
|
8 |
-
Because of this, I will also offer a NLP version of ZoinksNoob in this repo with the NLP loras baked in.
|
9 |
-
|
10 |
-
Older model below this line
|
11 |
-
___
|
12 |
ZoinksNoob is a Noob 1.0 personal merge based off the 0.75 version of ChromaXL Mix.
|
13 |
Use 3-ish CFG.
|
14 |
12 steps and up are good.
|
15 |
-
Credit where credit is due goes to Lodestone as I piggybacked off https://huggingface.co/lodestone-horizon/chromaxl-mix in the merge process. Utilizing output blocks from it for this merge.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
ZoinksNoob is a Noob 1.0 personal merge based off the 0.75 version of ChromaXL Mix.
|
2 |
Use 3-ish CFG.
|
3 |
12 steps and up are good.
|
4 |
+
Credit where credit is due goes to Lodestone as I piggybacked off https://huggingface.co/lodestone-horizon/chromaxl-mix in the merge process. Utilizing output blocks from it for this merge.
|
5 |
+
|
6 |
+
ZoinksNoobTest is an experimental version of ZoinksNoob with some Loras merged in that allow for a bit more lax prompting and simple natural language prompting. Can help with generalizing short 3-4 word phrase "tags" that don't explicilty exist in the e621/danbooru dataset.
|