Z-Image: Powerful and highly efficient image generation model with 6B parameters

Most of the people I know doing local AI prefer SDXL to Flux. Lots of people are still using SDXL, even today.

Flux has largely been met with a collective yawn.

The only thing Flux had going for it was photorealism and prompt adherence. But the skin and jaws of the humans it generated looked weird, it was difficult to fine tune, and the licensing was weird. Furthermore, Flux never had good aesthetics. It always felt plain.

Nobody doing anime or cartoons used Flux. SDXL continues to shine here. People doing photoreal kept using Midjourney.

> it was difficult to fine tune

Yep. It's pretty difficult to fine tune, mostly because it's a distilled model. You can fine tune it a little bit, but it will quickly collapse and start producing garbage, even though fundamentally it should have been an easier architecture to fine-tune compared to SDXL (since it uses the much more modern flow matching paradigm).

I think that's probably the reason why we never really got any good anime Flux models (at least not as good as they were for SDXL). You just don't have enough leeway to be able to train the model for long enough to make the model great for a domain it's currently suboptimal for without completely collapsing it.

> It's pretty difficult to fine tune, mostly because it's a distilled model.

What about being distilled makes it harder to fine-tune?

How much would it cost the community to pretrain something with a more modern architecture?

Assuming it was carefully done in stages (more compute) to make sure no mistakes are made?

I suppose we won't need to with the Chinese gifting so much open source recently?

AFAIK a big part of it is that they distilled the guidance into the model.

I'm going to simplify all of this a lot so please bear with me, but normally the equation to denoise an image would look something like this:

    pos = model(latent, positive_prompt_emb)
    neg = model(latent, negative_prompt_emb)
    next_latent = latent + dt * (neg + cfg_scale * (pos - neg))

So what this does - you trigger the model once with a negative prompt (which can be empty) to get the "starting point" for the prediction, and then you run the model again with a positive prompt to get the direction in which you want to go, and then you combine them.

So, for example, let's assume your positive prompt is "dog", and your negative prompt is empty. So triggering the model with your empty prompt with generate a "neutral" latent, and then you nudge it into the direction of your positive prompt, in the direction of a "dog". And you do this for 20 steps, and you get an image of a dog.

Now, for Flux the equation looks like this:

    next_latent = latent + dt * model(latent, positive_prompt_emb)

The guidance here was distilled into the model. It's cheaper to do inference with, but now we can't really train the model too much without destroying this embedded guidance (the model will just forget it and collapse).

There's also an issue of training dynamics. We don't know exactly how they trained their models, so it's impossible for us to jerry rig our training runs in a similar way. And if you don't match the original training dynamics when finetuning it also negatively affects the model.

So you might ask here - what if we just train the model for a really long time - will it be able to recover? And the answer is - yes, but at this point the most of the original model will essentially be overwritten. People actually did this for Flux Schnell, but you need way more resources to pull it off and the results can be disappointing: https://huggingface.co/lodestones/Chroma

> How much would it cost the community to pretrain something with a more modern architecture?

Quite a lot. Search for "Chroma" (which was a partial-ish retraining of Flux Schnell) or Pony (which was a partial-ish retraining of SDXL). You're probably looking at a cost of at least tens of thousands or even hundred of thousands of dollars. Even bigger SDXL community finetunes like bigASP cost thousands.

And it's not only the compute that's the issue. You also need a ton of data. You need a big dataset, with millions of images, and you need it cleaned, filtered, and labeled.

And of course you need someone who knows what they're doing. Training these state-of-art models takes quite a bit of skill, especially since a lot of it is pretty much a black art.

Thanks for the extended reply, very illuminating. So the core issue is how they distilled it, ie that they "baked in the offset" so to speak.

I did try Chroma and I was quite disappointed, what I got out looked nowhere near as good as what was advertised. Now I have a better understanding why.