KREA2 TE
1st version TE for Krea2 model with better prompt understanding and details
ZIMAGE TE
I and Triple Headed Monkey decided to publish another our uncensored TE that works for the Z Image Turbo and Z Image Base models. We have trained this using our own software. This is not the final version. At the moment, this gives more realism, more details in general, a more accurate picture, and a better prompt. Unfortunately, I don't have time to prepare comparative images right now due to some circumstances, so I'll just attach examples of images generated on my version of Z Image Turbo and Z Image Base (these versions are scheduled for release in the near future).
This version of TE differs from the previously published version THM_UNC_Z-image_TE - v1.0 | ZImage Text Encoder | Civitai
this TE for our Z Image Base fast model Lord Zolaris - ZIBv1.0 | ZImage Checkpoint | Civitai
And Z Image Turbo https://civarchive.com/models/2644869/solordz?modelVersionId=2969766
If you like our work and want support us please support us directly, since I am in a country which cannot get money from Civitai https://www.buymeacoffee.com/tripleheadg
our WF SoLordZ WF (upscale) - v1.0 | ZImage Workflows | Civitai
and https://civarchive.com/models/2662358?modelVersionId=2989530
Stay tuned for updates :)
Description
Krea2 TE V1.0
Improved prompt understanding and details
FAQ
Comments (4)
Can you hack the TE to always output whatever the prompt "amateur photograph from the 1990's" would output in addition to the input prompt? If you did that I think you'd be famous. Adding that prompt segment improves realism in Krea by a lot.
@ferrrett33 Could you explain in more detail what you mean?
@speach1sdef178 Try adding the sentence "Amateur photo from the 1990's" to anything done in Krea 2. It has a very interesting effect of making 80-90% of the images more realistic. The body shapes become far more natural and the tone is dialed down a bit. In about 10-20% of images it doesn't make it more realistic. I've tested this across numerous models and the effect is tremendous. You can also do different decades like 1980's, 1970's etc. but the further back you go the more decade specific it gets. It doesn't seem to work after the 2000's as well. I believe this is because if you put that prompt, whatever the AI creations were are forced to then better match patterns of actual photographs in its dataset. In other words, without that prompt Krea really doesn't do a "reality check" but that prompt forces it to somehow.
So I was thinking if there is a way to encode this either in the text encoder or the model weights to get the most realistic model ever.
PS: "photograph" worked a tiny bit better than "photo" in one test but I haven't tested that extensively, nor removing "amateur" or "90's" versus "1990's" versus like 1995.
before giving you 2 damn K buzz:
-Whats the process of creating this? (please provide full tecnical details)
-How much time required?


