Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 馃榿
Luscious lips, a canvas of hues untold, Creative colors, a tale to unfold.
Galena transformed, a feline grace, Reborn as a CAT in a Citrine embrace.
GalenaCAT IL v5 is cleaner than v4 but still has the nice soft shading and deep colors. It also has some of the creativity of CAT IL v10.
Update GalenaCAT_IL_v5 - (GalenaCAT_IL_v4:0.7) x (CAT_IL_v10:0.3)
Update GalenaCAT_IL_v4 - (GalenaCAT_NAI_v5:1.0) x (Unreleased_MoonLightCAT_Lora:0.3)
GalenaCAT Illustrious Base - Created by training a Lora on images I generated with Galena Redux v2 and Galena CAT [SD1.5] and then merging that Lora with Citron Anime Treasure at a factor of 1.0
Description
Created by training a Lora on images generated by Galena Redux v2 and Galena CAT [SD1.5] and then merging that Lora with NoobAI-XL Epsilon-pred 1.1-Version at a factor of 1.0. So its 100% GalenaCAT's strength on NoobAI.
FAQ
Comments (17)
i decided to try and make some images for the bounty, i ended up being impressed with the quality of the results and i made over 100 images, kkkkkk
Nice! I'm really glad you enjoyed it so much!
Update: New new NoobAI version! This one was made using the same merging technique as the main CAT Checkpoint and I think its easily 3 times more stable than the previous Galena CAT NAI.
.
New NoobAI version is fairly wild --> It makes nice colors but is kind of unstable lol. I have a more stable version but the colors aren't as nice.
So far as I can tell, NoobAI is a similar experience to Illustrious but its a little bit more better at following long prompts -- I'm experimenting with it and if you want to join me please let me know what you think! 馃榿
Out of curiosity did you try your Galena LoRA on RouWei? https://civitai.com/models/950531?modelVersionId=1330983
Oh this is an interesting Checkpoint! I'll test it out and post some images.
Here is a grid with 3 comparisons: https://civitai.com/images/69058895
Here are another 19 examples: https://civitai.com/posts/15344066
I like the results. Overall I give RouWei a thumbs up 馃憤
Sometimes the combination is great but not always. For example Marie is better with just Galena but Samus is better with Rouwei + Galena.
@CitronLegacy聽Ooh, very nice! Thanks for taking the time to test ^^
The latest noobai version I got some very dark and blurry image, any recommended configuration (cfg / samplers)?
Thats not good! Thanks for helping me test it. Can you try to copy everything (prompt, settings, configuration) from this image: https://civitai.com/images/69042398 ?
I used
CFG: 7
Steps: 30
Sampler: DPM++ 2M Karras
@CitronLegacy聽Thanks for your advice! It seems that the model is very sensitive to steps and samplers, it will get blurry and dark image when using my common use 20 steps + Euler A SGMUniform combination. and it seems that decrease the steps or change the sampler either will make generated image worse
for your information
@SLAPaper聽I dont have a way to test someone on a different computer so its really helpful to get feedback from someone's environment!
How does this other checkpoint work for you? https://civitai.com/models/131986 ? It was made differently than GalenaCAT so knowing if one works better than the other would help me figure out how to improve.
Hey I just posted a new version. I tested with 20 steps and Euler samplers and it seems great! (Example: https://civitai.com/images/69222275)
When you get a chance please let me know if it works better?
@CitronLegacy聽Just FYI, I generated a bunch of images on my computer, with CFG: 5, 20 steps, and samplers Euler / Euler a / DPM++ 2M using the newest checkpoint (NoobAI v5) and it worked just fine. Posted a few with the Toph lora.: https://civitai.com/posts/15393050
@chunchunmaru69聽Thank you very much! I really appreciate the test data!
Also that second Toph picture is perfection. It looks so professional! 馃ぉ
@CitronLegacy聽well, that being the case I've moved that picture to be the first on the post. 馃槀
After further testing on my end, the images only tend to get worse when we go to CFG 3 or lower, which is perfectly normal. To use super low CFG (1-2), go with the CFG++ samplers, those are working fine with this checkpoint even at only 18-20 steps.
@chunchunmaru69聽Thanks for all the thorough tests! Its cool that it works with only 18 steps. This made me want to try 10 steps: https://civitai.com/posts/15440406 The results are ok but look a little blurry. Its weird that low steps resulted in low resolution. 馃
@CitronLegacy聽I mean, 10 steps is really pushing the limits. The second image, seems pretty decent considering no upscaling.
There are loras for SDXL that make it work with even fewer steps. No idea if they work with Illustrious. https://civitai.com/models/350450?modelVersionId=391994