Intro

My Illustrious "Best of Both Worlds" semi-real/2.5d models

After downloading and testing numerous different semi-real Illustrious models, I found that many merges tend to lose color saturation, aesthetic detail, composition, and lighting the more they get closer to being semi-realistic. There are some really good semi-real models out there that achieve this to some effect, but also tend to be heavily biased towards specific subjects and concepts. After tinkering with various block ratios and text encoder balacing, these models are the result of my experimentation -- to achieve a "best of both" balance.

Overall, these models are less photorealistic than other semi-reals, but I am quite happy with what I've accomplished with them. Of course, it is not completely immune to overall Illustrious shortcomings -- realistic furry generation may be tricky and darker/low-lighting generations lean more towards 2.5D/Stylized influences since most photorealistic checkpoints are primarily composed of well-lit training data.

Nerds and creators can see additional merge details further below.

Merges:

The long-short: out of testing multiple batches of different merges with varying ratios, I've ended up with three that all scratch a specific itch.

Lyra

More stylized, less photorealism. Pretty artistic interpretation with painterly qualities. Lighting is dramatic and stylized. Facial features are simplified but expressive. Most intuitive and creative.

Echo

More realism and texture detail while maintaining the color palettes and lighting effects are photographic-like while maintaining anime aesthetics. More biased towards female body types despite OUT block tinkering. You'll see far less "accidental futas" when aiming for feminine men than most other semi-reals though.

Iris

Middle ground between Lyra and Echo. Closer to 3D than Lyra, but overall style maintains anime influence. Interestingly, environment rendering is really good on this one, as well as BL/Yaoi content (if that's your thing).

I used 99% identical prompts and seeds for each model's showcase, so you can compare and contrast them there.

Tags I usually use:

Positive

masterpiece, absurdres, best quality, newest, ambient lighting, very aesthetic, ultrahd, high contrast, realistic

Negative

worst quality, 2d, low quality, lowres, simple background, censored, mosaic censoring, censor, bad ???????, bad hands, ugly, artist signature, watermark

Generation Settings

CFG: 5-7

Steps: 36+ (Depending on complexity/elements)

Recommended Samplers/Schedulers:

(Starting with ones I've tested the most)

Euler A - Automatic | SGM Uniform(Helps blend problematic Loras) | Align Your Steps GITS(Leans towards 2.5d/2d)

DPM++ 2M - Karras | SGM Uniform | KL Optimal(Leans heavily towards photorealistic but very unrefined, recommend using with upscaler or post runs)

DPM++ SDE(Slowest, more realistic) - Karras | Polyexponential

DPM++ 2M SDE(A little faster than SDE while taking a minor hit in detail) - Karras | Polyexponential

Euler - Beta (Variable, leans towards realistic but is a bit of a wildcard)

These models are not 100% perfect but satisfy me for my first launching point in checkpoint merge experimentation for Illustrious. If I can improve on these with further refinements and lora merges, then I will continue to grow them if not spin them off into their own separate models.

Some nerd stuff for model creators:

Since the only difference between the models is block distribution and a handful of Loras, I'm counting the three merges as different versions of the same model. The checkpoints included in this are: knkLuminai(v1.0), Better Days Illustrious(v0.1), BoleroMix(v2.0), PornmasterPro(v1.8), CyberIllustrious(v1.6), and CyberIllustrious(v3.0).

The ways I weighted the merges are more or less like this:

Very minor differences in the IN blocks, with the non-realistic models taking up most of the weight (about 0.7-0.9). With Lyra, I did a jump up with the MID and then did a slow but gradual increase throughout the OUT blocks. Echo has an even more dramatic leap up. And Iris I wanted to keep some of the influence of the 2.5d in the OUT to see what it would look like, hence the slight dip. The CLIP for all three are weighted towards the 2D/2.5D models.

These are all sum merges too, with the exception of CyberIllustrious v3.0 which I did a differential merge to avoid overbaking similarities with v1.6. Nothing groundbreaking or revolutionary, but figured I would share if anyone was curious.