SiLAS - Feeltastic - The Real Beast - (16 Steps-Optimized)
A little special Present for you - A new and fantastic Photorealistic Checkpoint for creating beautiful Situations, People and everything else,
No Lora´s needet - (all Pictures above are pure without any Embedding or Lora) - so you get beautiful results.
.... have Fun !
I dont want to make Money with this so i dont ask for something, but if you want to help because you like my Model (which i hope) - let me know it... I would be thankful for any Buzz i got.... but its fine if you dont buzz - but please do write a friendly comment and... .... give a Thumb-Up !
Recommended Settings:
Res: 968*1408 with Kohya HRFix (i used Forge for every Image)
Sampler: Euler a - Exponential (Recommendet)
Steps: 16-35
CFG: 2.4 - 4 (just test it)
Description
FAQ
Comments (8)
It's a cool checkpoint, thanks.
However, the recommended options seem a little far-fetched to me. This is the turbo version of SDXL, so if you exceed cfg 1 too much and take too many steps, the images will be oversaturated. The recommended options would be cfg 1.5 and Steps:8, which will give the same result without the saturation. Euler a, on the other hand, is good.
Not that I want to criticize the work, thank you, but just to leave the information for those who pass by and see oversaturated images. :)
Hello, thank you for your comment. I'm an old fart, so I can handle criticism, yes, a very small amount of turbo has been embedded, but only minimally. If I actually choose the settings you recommend, the details i focused on disappears and the skin details are gone. Maybe we're both wrong and the truth lies somewhere in the middle. I think a CFG of 3 is actually really better—but 8 steps is loosing to much Details. 16-20 Steps make it crisp again, However, everyone has their own taste, and maybe everyone should just try it out a little. Personally, I get the most interesting music compositions when I experiment, and it's no different with Stable-Diffusion. I am still pretty sure that this Checkpoint has something really special...
@DiffusionShadow Hmm, you're right, after more testing, it's true that there is a loss of detail compared to your settings. And the saturation is no longer as significant as I had noticed, maybe it was a mistake on my part or some LORAs that had remained loaded without me knowing. However, yes, the ideal is somewhere in between, and a CFG of 3 as you suggest is really not bad. Looking closely at the images generated, I don't see any additional detail added after 20, or if so, only marginally.
Therefore, you're right, CFG 3 and Steps 20 seems ideal. I would also add tiling resizer X2 0.3 to complete the thing and it's perfect.
Great checkpoint in any case. Thanks. :)
@Illynir Thank you, have a nice Day and a nice Christmas-Time
i exatly want this type models(beautiful and realistic), but eyes and emotions barely show life.
Dmd make it worse, i think a normal model again with your experience would be better.
thx, nex week i will do a full version again... have a nice Time !
Hi, eventually take a look at the new Z-image model. For now, there's only the turbo version, but the base version will be released soon. It looks quite promising; it's much lighter than Flux and offers much higher quality than SDXL. I tested it in a workflow that refines SDXL images (I tested it with your checkpoint), and it's mind-blowing how well it refines and how compatible it is with SDXL images.
@davidOR Hi! Is there any chance you could share your refine workflow?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.