Lollipop is an extremely good 2.5D base model. She can easily handle a variety of styles, including but not limited to secondary, illustration, anime characters, etc. Lollipop is based my Flyingsnow model. It is a branch model based on my Flyingsnow model training + mix, Lollipop perfectly inherits the advantages of the Flyingsnow model, and on top of that, mix the Ghostmix middle layer and the 8th-11th output layer with 0.5 weight. This gives Lollipop the perfect inheritance of the strengths of Flyingsnow and Ghostmix, with the use of plugins to fine-tune the colors of the model.Anyway, I've always thought that a good large model should be versatile and not lean too much towards a certain style that is too fixed, as this can be left to LoRA. Therefore, when doing the final testing of this model, I focused more on its colors, composition, lines, backgrounds, characters, and the model's response to unstyled cue words such as Anime, Photographic, Digital Art, Fantasy Art, Neon Punk, Low Poly, Line Art, 3D Model Pixel Art, etc.Generation Parameters：Sampling Steps ：25-30Sampler method：Restart (regular images), DDIM (images with high quality textures. It will be slower than other samplers, but it doesn't matter if you run it online), PLMS (generates images with oil painting style)CFG scale：6-9Hires. fix Parameters：Step：15-20Denoising：0.1-0.3Upscaler：R-ESRGAN 4x+ Anime6B (Anime , Illustration), Nearest (try watercolor

OIA Lollipop

oiaLollipop_010.safetensors

0.10

Recommended Parameters:•CFG Scale: 6-9 •Clip skip: 2 (You can try 3-6, sometimes there may be unexpected surprises)•Sampler: Restar (or DPM++ 2M Karras / Euler / Euler a, these three are faster at producing images)•Sampling steps: 20-30 (Not too high, to avoid wasting processing time)•Size(SDXL): Nx1024, 1024xN (keep one side at 1024)•Size(SD1.5): 512x768, 768x512 (keep one side at 512)•Hires step: 15-20•Hires.fix: R-ESRGAN_4X+ (For images with normal style, the effect is better compared to 6B, other styles can be tried)•Denoising strength: 0.2-0.5 (High values will result in significant deviation from the original image)•VAE: Model already built-in Lollipop 0.10 Birthnote (Can be skipped):The creation of this model was inspired by my good friend Caramel Latte with Extra Milk, who said: "This model Flyingsnow goes really well with several of my LoRAs, when will you upload an older version of Flyingsnow to Lib." I am not one to repeat the same thing, so I came up with the idea of ​​creating a branch version. The advantages of Flyingsnow model: sensitive prompts, wide range of styles, and a Dark color bias. To complement Caramel Latte with Extra Milk's LoRA, I have been thinking about how to create this new version of the model. Initially, I thought about training a LoRA for fine-tuning, but eventually abandoned the idea because I didn't like the idea of integrating LoRA into a large model. So, I used my other model Dragonfly to run 500 style-consistent images and readjusted the Flyingsnow model. Perhaps due to perfectionism and OCD, when everything was adjusted, it felt like something was missing. At this time, I happened to see some pictures run by a friend's Ghostmix, with translucent colors and clothes texture, which is exactly what the Flyingsnow model needs. So, I said to Ghostinshell, let's merge your Ghostmix with Flyingsnow and see what effect it will have, and that's how the Lolliipop model was born. Caramel Latte with Extra Milk's LoRA (click to access):<a href="https://www.liblib.art/modelinfo/1e45441a09e148ac847b6def2f604619" target="_blank">Color Mosaic-caramel's Futuristic Illustration</a><a href="https://www.liblib.art/modelinfo/9343d772e1184f61b48504c993b7b7af" target="_blank"> &nbsp;｜ Caramel—Ramos Illustration|Color|European & American|Gorgeous</a> &nbsp;<a href="https://www.liblib.art/modelinfo/a799ecd3f9154167bd8a6ef4e8a70921" target="_blank">Caramel-Retro Textbook Watercolor Illustration</a> Thanks to: GhostInshell, Caramel Latte with Extra Milk

Files with SHA256: ede850f04823e801bfb1d64d6085a5b38f96bf56ee2c4177c083257a43ad6a16