THE TRUALITY ENGINE V2 - (VAE PRUNED)
(October 28th)
This model was re-designed / merged with a lot of attention to detail and versatile results.
If you had good results before, your results will be great now. Better saturation, more depths and greater sharpness when needed. It also work a lot better with most Loras. Embeddings should be used with CFG Scale 3.5 to 4.5
The least of us have tons of SDD-Space on our rigs and having 20 different models to achieve different results to realize your vision is not making things better and is also very time consuming.
My goal was to cherry pick models that deliver absolutely stunning and realistic photographs but also satisfy the 2.5 and fantasy lovers. It's also very "potent" with NSFW content. People who are creating series of images, i.E. Influencer Content, will find that this merge has a very high consistency. Hands and fingers are quite decent but not perfect.
HOW TO PROMPT?
I tried to use models where you have the ultimate freedom, if you are are used to slider from 0.5 to 1.5 you will do fine. If you are more the descriptive kind of person the model won't fight you. I have added a lot of examples with demos, no Loras been used, i used a few 2 TI's for influencer faces. Not to fool anyone but to show how well they work with this model.
SETTINGS
Steps: 25 to 35
CFG: 3.5 to 7 depending on how much saturation, freckles etc. you want. Wenn using TI's never go above 4
Best Size 512 x 768 or 768 x 1024 if you have a 3090 or higher
Best Samplers: DPM++ 2M SDE HEUN KARRAS or DPM++ 2M SDE KARRAS
HIRES FIX
UPSCALER: 4x_NMKD-Superscale-SP178000_G.pth
(Resize Steps 7 to 15, Denoise 0.17-0.33)
This upscaler provides a special layer of depth for photo realistic images.
REFINER
If you like the content of your favorite model more but would like to add some of the sharpness and lighting features to your output, add this model as refiner or upscale model. When using as refiner put at 0.85
VAE Pruned
LORAS
Most Loras will distort your results, you have to weigh them in, 0.25 to 0.65 seems to work best. Many... if not the most will mess with your models face because they were trained with a different face, sometimes you can fix it with clip skip and if that doesn't work just use roop or face swap and write your girls face over the trained face. Most Loras are designed for 512x512, when bodies or objects get strethed or appear twice then lower the output size.
WEIRD FACES
Especially from side angles and long distances you will keep running into weird faces, Hires Fix will 9/10 times fix this, if that still doesn't work use roop or face swap in your chain.
Please share your favorite results
