Personal merge based on Counterfeit 3.0 I've been using for a while (You can check more polished examples on my pixiv page). Sadly, after many add-ons, the recipe is lost in time like tears in the rain.
This model is focused on stylization and lora compatibility. Can't say exactly it's beginner friendly nor hard to use, you just have to prompt exactly what you want like outline, shading/lighting, and aesthetic/artist to keep consistency (by "artist" tags will give you some watermarks, so bear with it). It's pretty good for pastel/watercolor/oil painting without blowing everything up, try mixing and matching to find your preferred style!
Personal recommended sampler is ancestral Euler, but other samples will also give you nice results and different styles. If your image looks fried, just crank up the steps, if it's looking like a tumor, lower the steps, pretty intuitive:
CFG Scale: 5~10/ More to Less stylization and Softer to Sharper
Euler/2S a Karras/Restart: <25 steps / Lower denoising
K Samplers: 30~40 steps / Higher denoising
Upscalers: R-ESRGAN 4x+ Anime6B/4x-AnimeSharp
VAE: DogeVAE
These examples are not cherry picked or inpainted, you can get better results or worse, who knows...
Description
Same proposal as 1.0 ver. Here's what's new:
• Reworked for better precision at higher resolutions. Optimal at 512*768 +2x HiRes, you can even gen straight at 768*1024 granted the image is simple enough.
• Trained on cleaner art, well defined lineart/outline, should give you better precision on hands and clothing in general.
• Very sensitive to training, grabs on style very quickly, would recommend giving it a try.
• Trained on various western-ish artists like Cutesexyrobutts, Ratatat, IBUO, Shexyo, Nyantcha, Kajin, AraneeSama, a special assortment tagged as DogeBlend and many others, will have a complete list soon.
• This version is focused on improving the styles in general, ??????? and NSFW. Better clothing and scenery are plans for a future version.
• Even better prompt solving, inpainting and response to controlnet generations.
• Better precision on LoRA finer details and stylization.
• Baked VAE, but go for 840K if you want stronger coloring.
• The training was diluted to avoid fried images but still kinda strong, so go for CFG 5, should give you complete images even with 15 steps os samplers like Euler and go 30+ without degrading the generation.
• Focused on positive prompting, negatives as "worst quality, lowres" should be enough to keep the image straight, simple embbedings will also work for consistency, don't overdo on it, it'll compromise your generations.