With FLUX.2 Klein and Z Image Base (ZIB) out now, I’ve been doing some LoRA training tests.Honestly? My results so far have been pretty underwhelming. Here’s a quick breakdown of my experience.
There are three main factors that affect training results: **me**, the **training tools**, and the **base model itself**. On my side, that includes my training settings, my dataset, and how I evaluate results. Of course, issues with the tools or the base models can also hurt results—but I chose to start by blaming myself. So I kept tweaking configs and datasets. After about 70 test runs across Z Image Turbo (ZIT), Flux.2 Klein 4B/9B, and ZIB, I’m still not quite where I want to be.
I’m not a professional—everything here is just my personal, subjective take, and I could be completely wrong. It’s also very possible I messed things up myself somewhere along the way.
Alright, that’s enough preamble. Let’s get into it.
## I. Flux.2 Klein 4B
Long story short: I’ve given up on this one. I couldn't get anything even remotely decent out of it. It’s just not a good training partner.
## II. Flux.2 Klein 9B
Much better than the 4B version. I managed to get the facial features mostly right for character LoRAs, though there’s still plenty of room for improvement. I did notice some issues with facial warping and weird body proportions during generation. This might be fixable with better parameters or cleaner datasets, but I’ll save those experiments for later.
I rarely use Klein for T2I, but for I2I, it’s currently my top choice among the models I run locally.
## III. Z Image Turbo
I’ve been training on ZIT for a while, and it’s been a struggle. To be blunt, I’m not a fan of the ZIT "look"—the image quality feels a bit low-rent and lacks aesthetic appeal (though that might just be my poor T2I skills). That said, it‘s still my go-to for T2I among my local setups because of its speed.
ZIT is extremely **prompt-sensitive**. Almost every ZIT LoRA I’ve used (both my own and others’) can lose a lot of effectiveness depending on the prompt. I’ve tested this repeatedly, and even deliberately overcooking the LoRA doesn’t really solve it.
LoRAs can slightly improve image quality and style, but most of the time they end up being overridden by ZIT’s own look.
I used to think that once the base model came out, everything would magically improve. Looking at things now, ZIB and ZIT don’t seem very closely related at all.
## IV. Z Image Base
Personality, of the four models, ZIB is hands-down the best for training. It’s the easiest one to get a "successful" result from.
However, ZIB is in a bit of an awkward spot right now. It trains well, but it’s a nightmare to use for generation—it’s slow and the raw output quality is intimidating. We’ve all been spoiled by the speed of ZIT.
The dream was to train high-quality LoRAs on ZIB and use them on ZIT for the best of both worlds. Reality check: it doesn't work. I’ve tested this a lot. So far, I haven’t been able to get a ZIB-trained character LoRA to work properly on ZIT, even at strengths of **1.5–2.0**. Around 1.5, you can see clear character resemblance, but the overall result is still underwhelming. At 2.0, things often just fall apart.
Some people claim they’ve achieved very good results training ZIB LoRAs with **OneTrainer**, to the point where they can be used on ZIT at a strength of 1. I haven’t verified this myself, though—I haven’t used OneTrainer yet.
So the current situation with ZIB is:
it can produce good LoRAs—but those LoRAs don’t really have a place to shine.
Use them on ZIT, and the effect drops sharply.
Use them on ZIB itself, and ZIB’s T2I performance is just too weak.
So, ZIB is like the new SD1.5: great for training, but nobody wants to use the base model for generation. Its real value will only show up once we get a wave of fine-tuned community models. For now, all we can do is wait—and see whether ZIB derivatives take off like SD 1.5 once did. Unless, of course, a shiny new model comes along and kills the hype. In this fast-moving era, nothing stays king as long as SD1.5 or FLUX.1 did.
Note: My VG20 Rebecca character is the only LoRA I've released so far that includes versions for FLUX.1, ZIT, ZIB, and Klein 9B. It’s basically the "lab rat" for all these experiments.
Original link: https://sololo.xyz/model?id=586
