Z-Mania beta (Ultra realistic)
Still experimenting to better master the ZIT architecture.
I introduced into this merge a new type of LoRA (structural) essentially intended to act on the lower layers (blocks 0 to 10).
This second model has an almost exclusively realistic focus, can apply paint effects, but is not intended for illustrative or anime renderings. Your feedback is always welcome.
Z-Mania alpha (Precision Layer-Merging)
Z-Mania is a refined evolution of the ZIT architecture.
Alpha version Focus: High-end textures and color science via optimized Output Blocks (18-25).
Stability: Full ZIT compatibility (3072-dim) with original structural integrity.
Next: Breaking away from the ZIT "look" by overhauling Entry Blocks (00-10) for new spatial logic.
Description
same settings as the Alpha version
FAQ
Comments (33)
normal checkpoint or diffusion?
Normal, you mean AIO checkpoint? No, it's a model without VAE and text encoder.
@Adel_AI So I put this one in diffuser models and use a load diffuser model node
and a doubleclip and VAE
@serget2
Put this model in : ...\ComfyUI\models\diffusion_models
Download VAE : https://civitai.com/models/2168935?modelVersionId=2442479
put it in : ...\ComfyUI\models\vae
Download text encoder : https://civitai.com/models/2168935?modelVersionId=2442540
put it in : ...\ComfyUI\models\text_encoders
Is there any way to get this working with Forge Neo? The outputs are all messed up, and I've tried pretty much all the usual samplers, but nothing matches the "seeds_3" results in your pictures, any suggestions?
Sorry, as I'm not a Forge user, I won't be able to help.
@Adel_AI Ah, got it. No worries, thanks anyway <3
Have you tried any other checkpoint or just this one? Does another checkpoint work okay?
I haven't used this one yet but make sure you are on the Lumina tab (top of the page) and Euler A with DDIM works well.
@wideload Cool, that actually looks a lot better. It does seem a bit oversharpened on the details, but overall it looks neat—thanks for the tip!
@That_Red_Johny No problem. I made the same mistake myself at first.
Fantastic.
This work you've done is amazing! How well does the model support Chinese?
Thank you for your feedback. I haven't tested the Chinese prompting feature, you'll have to tell me about it.
How can I get the node file: DiTSelectiveMerger.py?
Article where you'll find the node download link in the attachment:
https://civitai.com/articles/23746/introducing-z-mania
Place the file in the custom_nodes directory and restart ComfyUI.
Can you add fp8 beta version?
Hello Adel,
I want your personal opinion, since you worked a lot with both ZIT and Flux Dev, if you had to choose a model, which one would you keep and why?
Salut Hamed, For a 6B model, ZIT is incredibly impressive compared to Flux1 (12B). Flux1 remains better for certain types of rendering, but overall ZIT surpasses it. I'm eager to work with Z Image's Base model; I think it will be better for fine-tuning and will likely become the undisputed leader in open source.
@Adel_AI What about Qwen? Chroma? I hope to see one day a model merged with the 3, Q-C-Z gathered together.
@vdalban127 The architectures of these models are not compatible, therefore the option of merging the 3 is excluded.
Can you add Beta in FP8 or Q8 GGUF version???
done
Hi,
I am experimenting with Z-Image and Flux models locally.
One limitation I constantly encounter is Arabic text rendering.
Current models can understand Arabic prompts reasonably well, but they struggle to generate readable Arabic words inside images.
Have you ever considered training a lightweight Arabic typography LoRA focused on:
Arabic calligraphy
Arabic signs and posters
Storefront text
Book covers
Common Arabic names and words
I believe many Arabic-speaking users would benefit from a LoRA specialized in generating readable Arabic text inside images.
Thank you for your amazing work.
Your comment touches on one of my main current concerns, since my work involves visual communication and I'm therefore often confronted with the use of Arabic and Latin scripts.
Yes, of course, it's possible to train a model on an Arabic font to obtain correct text renderings. But given the extreme variability of Arabic typography and calligraphy, if you train a model on "Kufi" it won't recognize "Diwani", the letter design is radically different.
Thanks for your feedback.
@Adel_AI Thank you for your quick response and for taking the time to explain the challenge of Arabic typography. I genuinely appreciate it.
I also wanted to say that I have a great deal of respect for your work. In my opinion, your position as the number one creator for Z-Image LoRAs is completely deserved. The quality, creativity, and visual refinement of your styles are truly impressive.
To be honest, one reason I contacted you is that I believe creators with your level of talent are rare. You have the ability to create artistic styles that are beautiful, sophisticated, and visually memorable.
As an Arab myself, and because I believe you may share a similar cultural background, I sometimes feel a little conflicted when I see highly skilled artists spending part of their talent on nude-oriented content. This is simply my personal view, not a criticism and certainly not a judgment.
At the end of the day, you are completely free to create whatever you want, and I fully respect that freedom. I just think your artistic abilities are valuable enough that they can stand on their own without needing that kind of content.
Regardless of our different perspectives on some subjects, I sincerely respect your work, your talent, and the professionalism you showed by responding so quickly.
Thank you again, and I wish you continued success.
@tamersalahibrahimelsayed
Thank you for your comment, perhaps a little too complimentary, but very touching.
Arabness is just one cultural layer among many, superimposed on my Berber, Mediterranean, and African identity... so it becomes difficult to claim a single identity, and I've never felt the need to.
@Adel_AI Any plan for next iteration? =)
I am working very sporadically on a V1, which requires dedicating much more time than I have available due to my professional commitments.
Thanks for your feedback
Q4 & Q3 version?
Sorry, I no longer work with these models, I’ve switched to KR2 and Ideogram4, which perform significantly better.
can we get this in int8 convrot ?