XL-TCD-Imitation Of Reality (IOR) SAT

After a long time, I honestly thought development on this model had reached its limit. The previous versions were not already delivering the realism that defines the IOR models, and I wasn't planning to continue working on it.

Then, during a series of experiments, I stumbled across a completely different set of training and optimization configurations. What started as a few tests quickly turned into something much more interesting. The new settings unlocked significantly better image coherence, stronger detail retention, improved lighting behavior, and more consistent character rendering, all while preserving the great realism that makes IOR series unique.

This update isn't just another incremental revision. It is the result of discovering techniques that allowed the model to push beyond what I originally believed was possible within the TCD architecture. Textures feel richer, environments have more depth, skin and material rendering are more convincing, and the overall visual quality is noticeably more refined without losing the model's signature dirty, lived-in aesthetic.

I had originally considered this project finished. Instead, it ended up surprising me. This version represents a new chapter for IOR TCD and serves as proof that sometimes the biggest improvements appear when you think there is nothing left to improve.

Thank you to everyone who continues to use, test, and support the model. Your feedback keeps pushing these experiments further than I ever expected.

There were significant changes in the recommended parameters; it is very important to try to adapt to them.

The VAE is BAKED

Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 10 denois 0.3 Hires CFG 1.4,

Width and height: 896*1152

CFG Scale: 1

Steps: 8

It's important to clarify that so far I've only achieved good results with EULER A, but it works with almost any type of scheduler: Euler A-Karras

Something quite important are the steps and CFG parameters; you can try between 6 and 12 steps and a CFG between 1.2 and 2. My sweet spot is with 10 steps and a CFG of 1.5

The most noteworthy thing is that I finally managed to recreate the aesthetics, the prompt understanding, and the appearance of a complete SDXL model, but with a good acceleration method, unlike the DMD, Lightning, and Turbo models. I think this one truly captures all the good aspects of SDXL.