Note: After a lot of testing, I think you can try the AYS schedulers, Gits, 11 and 32I want to start by saying that this model is quite experimental, remembering that it is an acceleration method that will work in 8 steps so please don't be too harsh in your judgment. Even so, I think it has given me very surprising results in many things.To begin with, what is this? It's an SDXL model based on the TCD acceleration system, a system that, unlike Turbo, Lightning, or DMD, delivers results much closer to a true base SDXL. If you want to know more about this, I've included two very informative articles.<a target="_blank" rel="ugc" href="https://civarchive.com/articles/29730/sdxl-acceleration-methods-dmd-vs-tcd-vs-lightning-vs-turbo">https://civarchive.com/articles/29730/sdxl-acceleration-methods-dmd-vs-tcd-vs-lightning-vs-turbo</a><a target="_blank" rel="ugc" href="https://civarchive.com/articles/29775/tcd-for-sdxl-the-acceleration-method-that-feels-surprisingly-human">https://civarchive.com/articles/29775/tcd-for-sdxl-the-acceleration-method-that-feels-surprisingly-human</a>The real advantage of using TCD is that, unlike DMD, for example, which only works with the LCM Sampler (which many complain about for its plasticky textures), here you can use a wide variety of Sampler-Schedule combinations.This model is experimental, as I said, since the TCD method is designed to run in Comfy with a specific sampler (TCD), but here I achieved a hybrid version so that it can be used not only in Comfy but also in Forge (as well as other methods, but for now I will only refer to those two).The parameters to be used are very important, so I recommend the following:<ul><li>The VAE is BAKED</li><li>Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 10 denois 0.3 Hires CFG 1, i changed only the Hires Sampling to Euler A but the other parameter stay the same,</li><li>Width and height: 896*1152</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>The sampling methods I recommend, although you can experiment with several, are: , DPM++SDE-Karras, DPM++SDE-SGM Uniform, DPM++SDE-Beta, Euler A-Beta, Euler Beta, DPM++SDE-AYS Gits</li></ul>The model works very well with CFG at 1, so there's no need to use negatives.What do I mean by saying that different samplers and schedules can be used? Judge for yourself.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f2d24fae-7eae-4be9-af46-78904b511ae9/original=true/f2d24fae-7eae-4be9-af46-78904b511ae9.jpeg" />The prompt used for that image was: Vintage photograph of a Chinese woman in triad clothing in front of an arcade in YokohamaI really hope you can try it, give me your feedback, and help me improve. Keep in mind that it's one of the few SDXL models with a TCD system. Thanks.

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.<ul><li>The VAE is BAKED</li><li>Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 10 denois 0.3 Hires CFG 1.4, </li><li>Width and height: 896*1152</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>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</li><li>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</li></ul>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.

Files with SHA256: 41c153ea07eb7bc3c29c3dec0b4b7eb9db2fd54ed4f88cd04b56dc158eb7a82d