The main goals of this checkpoint are, by order of importance, those listed:

• cartoon/anime styles

• LoRA fidelity

• rich compositions

• prompt fidelity

• higher resolutions

• flexibility

Despite being focused on 2D and flat styles, this checkpoint still has a pretty good range for 3D and realism with a little push, accurate even for real person LoRAs. Baked characters and styles are pretty consistent and PONY scores are less relevant. There are two versions of this model, one with pretty tight fitting, very consistent and on rails and one with loose fitting for more flexibility on shapes and textures, still the recommended methods are the same on both, so pick accordingly.

These checkpoints are finetuned for generating with upscaling, that means they have unconventional ETA/Noise/Denoising for proper latent upscaling and the optimal settings will really depend on your type of image and resolution, for resolutions around the 1024 you can go for high denoising like 0.7 and for 1280 ranges you can go for 0.6, the strenght of the scheduler will also be influential on that, you can start at "Normal" and go to "Karras/SGM uniform" if you feel like you need more denoising. One method for knowing you've reached your optimal settings is to generate some stuff with consistent artist watermark or something similar and check the range where the text will be most legible.

Try to familiarize yourself with the sampler parameters settings, they were also important for the testing of this model, mainly generating on different ETAs, sigma min/maxing will also help you shape your generations however you want them.

Check the recommended VAE, it is key if you want to reach the right results. Also a good read on the model page might also troubleshoot and save you some time.

Recommended Samplers: Euler/Euler A/DDPM

Recommended CFG: 5-7

Recommended Upscaler: Latent/Lanczos/Tiled upscalers

Recommended Denoising: 0.6~0.7

such pone