LoRAs are back at FLORA

LoRA training and generation are back in FLORA, now on Flux 2 and Krea 2. We pulled them a year ago. This is what changed.

For the first few years of FLORA, the biggest hurdle to adoption was consistency. Style consistency, character consistency, brand consistency. Anyone working with AI has felt this, but it was particularly painful for us, and it came up on most customer calls. It was the line between FLORA being a toy and FLORA being a real professional tool.

What a LoRA actually gave us

LoRA is a lightweight technique for fine tuning models. Think of it as adding a few layers on top of a model to help train it for a specific concept or look. Take a few images of a certain theme and it could replicate that theme seriously well. It felt like removing the shackles of the training data bias, and the results blew us and our customers away.

There was a catch. Several, really:

• Training was slow. SDXL took 15 minutes to train a single LoRA.

• They were big. Almost 300 MB.

• They needed a lot of setup. A set of many images, labeling for each one, all zipped up a particular way with particular naming conventions.

• They were finicky. Steps, trigger phrases, scale. All of it needed experimentation to get right.

Why we tried to hide all of that

When we first brought LoRAs to FLORA (funny coincidence on the naming, I know), we wanted to be maximalist about it: the power of LoRAs without the pain. The plan was to abstract everything away into Styles, which would train LoRAs across as many models on the platform as possible and fill in the gaps with prompt-based or image-based style manipulation. Styles would be universal. We had a vision of a library that could be applied to any model on the canvas.

That didn't work as well as we had hoped.

One lesson was that over-abstraction can be more confusing from a UX perspective, especially when we couldn't get the models to do what we wanted. The other lesson was about who we were actually building for. Expert AI users already knew what a LoRA was, could leverage the technology, and understood its limitations. Once we abstracted it into Styles and promised simplicity, 15+ minutes and $10+ stopped reading as an understood limitation and started reading as a raised eyebrow.

The goalpost moved.

Nano Banana, and the decision to pull Styles

In August 2025, Google released Nano Banana. Its level of prompt understanding let users semantically edit images at really high fidelity, and it got us roughly 80% of the way to style adherence on its own. That threw off the effort-to-reward scale completely. Next to it, LoRAs felt clunky, expensive and difficult to understand. We pulled Styles from the platform in late 2025.

Why they're back

A few months later we started hearing about LoRAs again, and from a specific group. For professional users, the training data behind the base models often didn't have enough context on a custom style or a particular face, and a LoRA was the right way to fill that gap. For those users the effort-to-reward math still worked. This time we wanted to be eyes-wide-open about who the feature is for.

LoRA upload and generation are now live in FLORA for Flux 2 and Krea 2. Bring your best custom styles, characters and brands, and your outputs stay exactly consistent.

Where they excel is styles that are difficult to verbalize and that aren't in the training set of the image models. It's a power tool, not a shortcut.

An example: acupuncture diagrams

I find the style of acupuncture diagrams interesting, so I found a set of 10 and trained a Krea 2 LoRA off them.

That took about 10 minutes to train. You'll find it in the right parameter bar on an image node with a LoRA model selected.

From there I applied the style to some other subjects.

Those aren't real Chinese characters. But the pen-and-ink style, plus the pointers laid out as if it were a diagram, line up quite well. Getting there with a typical image model would take detailed prompting and a lot of effort. That's the power of a LoRA.

A second style, same subjects

Here's another style I trained in FLORA:

And the outputs, running the same two subjects, plus an NVIDIA graphics card for fun.

Take it for a spin

LoRAs are on Flux 2 and Krea 2 today. Open a canvas, drop an image node, pick a LoRA model, and go make something.