
An SDXL model that generates anime waifus of consistent style and quality with minimal prompting, it can also do males and humanoids in general. It's fairly capable in both sfw fashion and nsfw concepts while being well behaved. Works with most concept and character lora, pony included, stye lora tend to have less impact. Baked VAE.
Prompting tips
Pony scoring and quality prompts are unnecessary.
Don't expect photorealism but using "photorealistic" or "realistic" at the beginning of the prompt will give a different, more 3d effect, while "anime artwork" will ensure more 2d faces. Using "3d" in the negative prompt ensures a flatter anime effect, when combined with the above this can give some nice results. The impact of this becomes less distinct with more complex prompts, though you can play around with the weights.
All merge testing was done using Negative Dynamics XL as the negative prompt, without it expect mixed results and often cartoonish body proportions, sometimes this is great.
Use whatever negative prompts you find give you results but in general anything that isn't necessary is just limiting output to disadvantage. Words like ugly, deformed, monochrome etc. usually don't have the expected effect or just make output worse, useful negative prompts are things like detached sleeves, lamps, shoes, 3d etc. and used when needed.
My general settings
DPM++2M SGM Uniform (other samplers work fine)
23 steps (the sweet spot for DPM++2M)
896 * 1152
7 CFG, lower can work well
Hires. fix:
4xRealisticrescaler_1000000G (I find this affects colour the least but there are other options)
18 steps (it's enough)
0.3 Denoising (higher can fix more imperfections but also strays further from the source)
1.75 Upscaling (I find this is about the limit for source accuracy with XL)
描述:
训练词语:
名称: envokenAnimeXLDelta_v10.safetensors
大小 (KB): 6938817
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success