Returning after a pause.
I left Civitai for a short time due to personal circumstances. However, during my absence, together with Triple Headed Monkey, we did not stop working and conducted several major projects.
Today I decided to publish one of the very first, but at the same time strong versions of our joint development — the improved ZIT base. Key changes compared to what was before: increased detail, reduced background blurring (less blurring) and improved realism.
Thanks for waiting. We are back and ready to continue working.
We use our training data and our soft for training.
!!! For best result use this TE THM_UNC_Z-image_TE - v1.0 | ZImage Text Encoder | Civitai
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
for relism use euler/normal, for art sometimes better result give euler_a/normal
All images in the gallery just raw generation w/o any tricks or upscale
Description
Returning after a pause.
I left Civitai for a short time due to personal circumstances. However, during my absence, together with Triple Headed Monkey, we did not stop working and conducted several major projects.
Today I decided to publish one of the very first, but at the same time strong versions of our joint development — the improved ZIT base. Key changes compared to what was before: increased detail, reduced background blurring (less blurring) and improved realism.
Thanks for waiting. We are back and ready to continue working.
We use our training data and our soft for training.
!!! For best result use this TE THM_UNC_Z-image_TE - v1.0 | ZImage Text Encoder | Civitai
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
for relism use euler/normal, for art sometimes better result give euler_a/normal
All images in the gallery just raw generation w/o any tricks or upscale
FAQ
Comments (14)
Welcome back! 😉
hi, what's the difference between the different versions of 3-4b?
normal vs josiefied vs z-engineer-vs fedric?
Hi! "Normal" base TE model Qwen3-4b good also good but with version by fedric you get more details and his version TE better following the prompt
We've just uploaded our own Text Encoder model. If you'd like to compare that and give us your feedback on it, that would be greatly appreciated.
Can you please explain why all your examples have neg prompt for distilled zit with cfg 1?
Hi! :) This negative prompt does not affect anything because the model works on CFG=1, it remains in my WF from the Z Image Base. ComfyUI ignores that negative prompt anyway. So just ignore it :)
Review: 5-24-2026
This one does good in terms of realism. However, it doesn't do NSFW. Probably needs the xxx lora for that. The recommended fedric 3-4b is not as good as the josified. Day and night difference for my personal use. I'm interested to see where this one goes.
Thanks for the feedback. This is just the beginning. We have accumulated a lot of material and we have done a lot of work in various fields, mostly rebuilding the model from scratch. We will publish it gradually, and this is not the final stage yet. The final is just ahead
@speach1sdef178 thank you for the reply. i wasn't too harsh in the review as i read your situation and you probably missed a lot of stuff that has developed. It was only recently that the true "realism" made a breakthrough and that was with people mixing each other's checkpoints. Do you plan to develop an all-in-one or plan to do the sdxl way that has a strong base and allow loras to be used to allow the force use of the lora concept? So far in my experience with most of the checkpoints is that some do better than others in terms of prompt adherence. And some loras that allow a certain pose messes up the checkpoint. But so far zimage isn't really good at keyword concepts yet. it requires explaining in detail what one wants to see and even then its not guranteed.
@Melodic_Possible_582589 The Z Image is a specific model and it is quite difficult to manage for those who are used to working with previous generation models such as SDXL and requires some skill to create a prompt, but this model is worth paying attention to because it can do incredible things that other models simply are not capable of. We are working on all possible options, and I think the finale of our work will surprise many. I can't reveal any more information yet.
@Melodic_Possible_582589 Our plan is to build a strong base. We have trained a lot of topics, but we haven't gone into advanced poses etc. or anything truly XXX. We've left it soley at the anatomy on a professional level. And yes if you would like to do more with the model it will need custom lora on top.
@Triple_Headed_Monkey the problem of custom loras is they are trained on a vanilla ZIT and mess outputs of finetuned checkpoints.
@mphobbit Of course. If you want the best compatibility the lora should probably be trained on the same model.
@mphobbit I remember reading that if people trained the lora on z image base the problem would go away or lessen. But it still depends on the checkpoint as well. There are still a few checkpoints that get ruined when the lora is at a weight of 1.0.