What is this? This is a merge of many illustrious models, with modified block ratios at each step, to make a well balanced and clean looking checkpoint for semi-realism, anime, and furry generations. The goal is for this to be a checkpoint that requires minimal LoRAs to make good looking images with a variety of prompts.

Usage Tips: To get smooth and colorful results I recommend starting with Euler A, a CFG of 3.0-6.0 and 22-30 steps, and adjusting as needed. For photorealism you may want a lower CFG and higher step count, however :)

The model is somewhat biased towards generating women even if you have "1boy" in the prompt, so if you want to generate men you may need to weight the tag or use multiple tags such as "1boy, male"

All example images are generated using a ComfyUI workflow which involves a high-res fix step with 0.30 to 0.50 denoise and 25 additional steps (50-55 in total). For the upscaling I apply gaussian blur with 1.0 radius and 0.5 sigma (softening the image can cause SR models to add a bit of artificial detail and also helps with resampling) and then upscale with the following MoSR model: https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr

Versions:

V1.5: Enhances contrast, saturation and in some cases prompt comprehension. This is not always desirable over the look of v1.0 however.

V2.0: I created anime, furry, and realistic finetunes based on v1.5 and then repeatedly merged the 3 versions together with varied block ratios to preserve some details while overriding others. The result should have a generally improved understanding of details and a more diverse range of themes & creativity.

V3.0: Merged 14 additional times including Illustrious XL V2.0, using DARE-TIES and slerp method. I also added Smooth Detailed Booster v3 and SPO-SDXL-locon to the weaker model at each step but only if it improved the results. You may want a lower CFG for this version as it has naturally higher contrast than previous versions.

V4.0: This version might be considered more of a sidegrade of v3.0 as it doesn't always look better. But it (hopefully) knows some characters better and has a larger range of creativity & styles than v3.0. Also, the color is toned down a bit. Use simple prompts and try different seeds as results can be highly varied.

Flash versions are merged with SDXL-Flash to reduce the number of steps needed for inference, at the cost of results being slightly lower quality on average.