This is the initial alpha release of a new AIO checkpoint,
intended to take Z Image Turbo to new places...
Announcing JoZiMagic:
an AIO checkpoint using ZImage Turbo with a different clip encoder and VAE that should change up the results you get... for the better.
It's an AIO checkpoint consisting of:
Normal bf16 release of Z Image Turbo
'better' Clip Encoder by replacing the stock Qwen3 4b encoder with a "Josiefied Gabliterated" version (hereafter known as Josie) ** in testing, this encoder produced very different but consistent (and often superior) results given the same seed/prompt/etc. as the stock encoder clip model.
'different' VAE than the stock flux VAE, currently the G-REPA/Self-Attention-W2048-3B-Res256-VAEFLUX-Repa0.5-Depth8-Dinov2-B_100000 VAE ** This might later be replaced with a different VAE, such as Anzhc/Z-Image_Anime_VAE, depending on what seems to perform best...
Yes, you can put these together yourself from the parts... and it's a big checkpoint (33gb) so putting together some quantized versions is a good idea... but you gotta start somewhere. By making this an AIO, it encourages people to try it out. And it lays out a name for this combo of things...
Load using the normal 'load checkpoint' node in ComfyUI, and add the usual nodes: Text Encoder for the prompt, a Sampler, a VAE Decode of the Latents and Save Image.
I recommend using the Zimage Text Encoder nodes current found inside https://github.com/fblissjr/ComfyUI-QwenImageWanBridge as those encoders give you MUCH more control over what is sent to Josie.
I'll add some sample images and workflow in a bit...
Thank you to the team behind Z Image Turbo: TongyiLab
Thank you to Gökdeniz Gülmez for the J.O.S.I.E project, which retrains a variety of models to behave differently.
Thank you to the G-Repa folks for the VAE
And a huge Thank You to the people on the Banodoco discord, including Kijai, Pom, and all of the other regulars there in the wan and zimage channels.
Description
AIO merge of stock Zturbo with a different (Josied) Qwen3 encoder, and an alternative VAE.
Loadable by the normal ComfyUI checkpoint method...
NON quanted for now, so full sized...
FAQ
Comments (20)
Not working on weak machines :(
Sorry, first draft, full models, I will release some quant-ed versions though.
Well the creator did say it was a non-quantized version. That should have already given you the first clue that it's not gonna work on all machines just yet.
30 gbs is crazy
@Thatonecreator fp16 Zimage, full josied qwen3. Runs JUST fine on my 16gb 5060ti. Not crazy.
@scruffynerf Yea cuz your on 16gb which not everyone has lmao. My guy the fp16 z-image model is less than that
@Thatonecreator
1) already said smaller versions are coming, working on figuring out the best 'combo' that doesn't descend into hell imagery... q3 did that... q4 not so much, looking at q4 or q5 right now.
My goal for 'JoZiMagic Trinket' (the small version) is around 7gb in total (that's model/clip/vae AIO)
2) THIS checkpoint includes the fp16, it's half of that 30gb size, so perhaps you don't understand what this is. Read the docs.
@Thatonecreator Runs just fine on my 24gb 3090, but that is just me. Although back then when I used to have a 3060, I think it would run fine anyway tbph.
@rommix0 Bruh
@scruffynerf Sure it does buddy🤣
It also works on my 4070 with 12GB. Comfy seems to take care of it, maybe only loads the model partly, which does not make the results bad.
@johnjohnson That seems to be what's happening, because even with something as powerful as the 3090 I thought the model wouldn't load. Maybe having 48GB of regular RAM helped.
@permitted you're an idiot, I'm doing multiple projects.
suggested sampler and scheduler?
you can pretty much use it with anything zmagic works with... haven't found any issues at all.
@permitted I personally recommend dpmpp_sde and bongtangent combo, but plenty of others work
That's not called Josied, that is called Josiefied! And it doesn't already change the output
yes it does. You don't know what you're talking about. I can post side by side compares, radically different.
Why Self-Attention-W2048-3B-Res256-VAEFLUX-Repa0.5-Depth8-Dinov2-B_100000 VAE? What does it do better?
it was my choice at the time, today I'd probably do UltraVAE. I need to release a full release, but life got in the way, as did Zimage Base, Klein, and LTX 2.3