|
--- |
|
license: cc |
|
--- |
|
Brand-new checkpoints series where a model is merged and upgrade with only my resources (own LoRAs and datasets). Targeting a sweet style, colorful, pleasant and cute. |
|
|
|
**Entrance :** |
|
|
|
First paths, the base model is 0.7xMIX_11 + 0.15xMIX_8 + 0.15xMIX_12, 19 homemade LoRAs were added to the model with high strength, values had been tweaked to keep at the end the best output. Kind of a half-trained checkpoint. |
|
|
|
Warning : during testing, I found that 2 of the negative embeddings I used in the past were hardly impacting the style, by getting rid of them, I could have access to a better and more colorful style. |
|
|
|
4 Groups were made to ease the tweaking, see screenshot : |
|
|
|
(It is like merging 19 models that are having one LoRA attached to it with high strength) |
|
|
|
|
|
**Settings recommended :** |
|
|
|
Sampler : Euler a + Automatic/uniform |
|
|
|
Steps : 27-30 |
|
|
|
CFG : 7-8 |
|
|
|
Hires Fix : R-ESRGAN 4x+ Anime6B or 4x-UltraSharp + Denoising strength 0.5-0.7 and 20 Steps (upscale x1) |
|
|
|
Upscale : img2img x1.5 - Denoising strength around 0.5 (lower if good input, higher if more corrections needed) |
|
|
|
Edit : Controlnet, openpose, depth (midas), Krita AI diffusion -> use Alternative - Euler A Preset, live mode to add or remove things (select area with selection tools) |
|
|
|
Enhance : after img2img and other editing, -> Extras tab, upscale x2 with R-ESRGAN 4x+ Anime6B or 4x-UltraSharp ; |
|
|
|
R-ESRGAN 4x+ Anime6B -> Better colors and anime looking |
|
|
|
4x-UltraSharp -> More realistic, more details |
|
|
|
Other Working Samplers : |
|
|
|
DPM++ 2M, DDIM, Restart, UniPC -> bad with Hires Fix |
|
|
|
DPM++ 3M SDE (expo/polyexpo) / DPM++ 2S a / DPM++ 2M SDE -> compatible Hires Fix |
|
|
|
|
|
**Comparison with CSA series :** |
|
|
|
Style : **+** |
|
|
|
Stability : **-** (removal of negative embeddings and high strength LoRAs) |
|
|
|
Creativity : **+=** (do more easily heterochromia, gradient hair and monster girls) |
|
|
|
Prompt listening : **+=** (trained with booru tags, better identification of some concepts like thin legs, shiny skin...) |
|
|
|
Render time : **=** |
|
|
|
Aging : **Younger** (could add chibi in negative prompt to avoid big heads) |
|
|
|
Hope you will have fun with this heavenly checkpoint ! |
|
|
|
Last but not least, since there is a strong influence of WingedCloud and Blue-senpai artists, the model is kinda naughty so don't hesitate to put nsfw in negative prompt ;). |