Any updates on the Base and Edit models?
A rough time estimate will be much appreciated! Like another month or 2 months? end of 2026? Q2 or Q3 of 2026?
Just to clarify : I'm like you :)
But, looking at the gallery, there is an image where you can read :
The image is awesome! But what to infer?
I'm very excited and looking forward to the release of Z Image Edit. Z Image Turbo is very good, it has my full support and I'm very happy with Z Image.
不会那么晚，最新消息有50%概率在中国的春节前后发布，也就是2月17日前后。也有50%概率本月内发布，首先发布的会是omni base和zimage，edit会在更晚时间推出
2028 Q4
Honestly, just download flux small 9b. The model is better, and there is already a basic model and distilled version available. You have a perfect reference method that works incredibly well, and LORA support will be available today or tomorrow. No one wants to wait any longer for news from Zimage. By the way, the image editing is incredibly good, and I don't think zimage can compete with that.
Honestly, just download flux small 9b. The model is better, and there is already a basic model and distilled version available. You have a perfect reference method that works incredibly well, and LORA support will be available today or tomorrow. No one wants to wait any longer for news from Zimage. By the way, the image editing is incredibly good, and I don't think zimage can compete with that.
I really not agree. Flux2-9b is very bad compared to Z-image-turbo. Yes it can make edit (that's better than today Z-Image), and the image quality is good (as ZIT is), but on T2I, the prompt following is very bad compared to ZIT, plus it's very bad in anatomy (3 arms or legs, doesn't follow prompt, 6 fingers, ....; and if you try 2 or more persons, you will laugh).
I really have the feeling to be returned 1 or more year back. The only good thing, from my point of view, is it's editing capacity which is very quick with the klein version.
And I'm not speaking of the speed of lora training (10x slower than ZIT on a 8Gb Vcard). <- note that I don't know if this is the "turbo" in ZIT which make it 10x faster, maybe it will be the same as FLUX2-9b when the base model will be released.
@ZetofI have no idea which model you tested, but so far I haven't used a better or more realistic model than fluxklein9b. Maybe you have the wrong settings or just a bad PC, no offense. I've created such perfect LORAs that they're almost indistinguishable from reality, and fluxklein9b requires only a few steps during training and learns anatomy so perfectly. Wait until good LORAs appear if you can't create them yourself, then you'll see how great the model is. Test it thoroughly and play around with the settings. The distilled model produces photos in such good quality in seconds that zimage can't compete, believe me. And on top of that, there's the perfect reference photo, which also works perfectly. I really don't know how anyone can badmouth this model; it's a complete mystery to me. I strongly suspect that you did something wrong or looked at some comparisons on the internet before you properly tested everything yourself. Don't use any ready-made workflows; build something yourself and test it until you get it perfect.
@ZetofI have no idea which model you tested, but so far I haven't used a better or more realistic model than fluxklein9b. Maybe you have the wrong settings or just a bad PC, no offense. I've created such perfect LORAs that they're almost indistinguishable from reality, and fluxklein9b requires only a few steps during training and learns anatomy so perfectly. Wait until good LORAs appear if you can't create them yourself, then you'll see how great the model is. Test it thoroughly and play around with the settings. The distilled model produces photos in such good quality in seconds that zimage can't compete, believe me. And on top of that, there's the perfect reference photo, which also works perfectly. I really don't know how anyone can badmouth this model; it's a complete mystery to me. I strongly suspect that you did something wrong or looked at some comparisons on the internet before you properly tested everything yourself. Don't use any ready-made workflows; build something yourself and test it until you get it perfect.
Hi, thanks.
You're certainly right. But I made more try, and I couldn't achieve the ZIT quality/prompt adherence. And I tried to make a second lora (the first is really good with 3000 steps/66 images => 45 epochs), and now it takes "only" 2 times more than ZIT to train (it was 10x for the first). So I don't know what's changed. The computer where I train is not the same where I infer. (training on a notebook RTX4060/8Gb with 32Gb ram; testing on desktop RTX5060ti/16Gb with 64Gb ram).
But For my last tests, I found the quality not so good, with noticeable anatomy disgraces (even with the lora find on CIVITAI to correct it), and the realism is sometimes not here.
Should make more tests, but for now I'm staying on ZIT, which is better from my point of view.
Anyway, the editing capacity is awesome !
Flux2-klein-9b is not a good alternative if you compare the license with Z-Image. Flux2-klein-4b has a good license, but it's not very good. I just hope that Z-Image Edit and Base will have a similar license as Turbo.