𝑹𝑬𝑨𝑳 𝘔𝘐𝘕𝘛⋆ is an expanding collection of meticulously crafted experimental models, each consisting of careful block-level merging of a large number of existing checkpoints - some well known and some very obscure, - as well as several checkpoints trained on custom, purpose-built datasets - which aims to provide:

Realism, particularly in human subjects.
An interesting uniqueness in its choices.
A subtle human imperfection not found in a lot of models.
Flexibility in adapting to most LoRAs.
High dynamic range for preservation of details.
Consistency in its particular abilities.

Recommended settings:

Sampling: Most methods seem to work well, with DPM++ 3M SDE Exponential (35+ steps) being a personal favorite for ALPHA, and DDIM for THEOS.
CFG: 3.5-7 (Dynamic Thresholding highly recommended for higher values)
Steps: Surprisingly good results have been observed with as few as 10 steps.
VAE: Baked in. (ALPHA: difConsistency RAW - THEOS: vae-ft-mse-840000-ema-pruned)

Note:

The development of this line of models is an ongoing long-term work in progress.

I do not own a GPU and I'm already struggling a lot financially, so development will happen on the rare occasion that I can afford food AND compute.

If you like my models (or don't), please don't hesitate to comment or reach out for any reason! Your reviews and feedback means a lot and keeps me motivated! ❤️

Model comparison:

Each model has its own distinct strengths and weaknesses. You are encouraged to try both, to see which one you prefer. During development, I wrote these notes on the difference between ALPHA and THEOS:

ALPHA model:

More flexible
Less people focused
Bit better backgrounds
Bit more realistic
Bit more vibrant
Less CFG forgiving
Better human "imperfections"
More contrasty and gritty
Subjects less likely to pose

THEOS model:

Really likes to generate women
More photoshooty
More horny
Better fine details
Harder to prompt
"Cleaner" more polished look
Smoother skin
Less artistic/creative

This flowchart might help you decide which model to try first: