Introducing the Flux.1 Krea [dev] On fal

Introducing the Flux.1 Krea [dev] On fal

Recently through a partnership with Black Forest Labs, Krea has a released a open weights fine-tuned version of FLUX.1, FLUX.1 Krea [dev]. fal has shipped API endpoints for inference and training at blazing fal speeds.

Exploring the Flux.1 Krea [dev]

When FLUX.1 was released last year it was a monumental step forward in the image generation space. A year after its release, it continues to be one of the most popular image generation models.

The FLUX.1 Krea [dev] checkpoint builds on the great base that is FLUX.1 and addresses weakness in the original FLUX.1 model. Here we compare two images with the same prompt.

On the left we have Flux.1 [dev] and on the right is Flux.1 Krea [dev]

They are both wonderful images. We are truly living in magical times, where by with mere words we can summon so much beauty, but FLUX.1 has some common attributes which can be a distraction.

If you look closely, you can see the face is a shiny. The chin has a dimple. Regardless of the seed, Flux.1 tends to generate men and women that looks very similar.

FLUX.1 Krea [dev] on the other hand, can generate a variety of identities from similar prompts.

Here we are using the simple prompt "woman':

The left four images are Flux.1 [dev] and the right four images are Flux.1 Krea [dev]

The left four images are all generated with FLUX.1 [dev]. Notice how the identity of the woman is very similar. each image has a caucasian woman with brown hair, a sharp jaw line, rosy cheeks, mouth slightly open and of course, the Flux.1 chin dimple.

On right, we have four images with FLUX.1 Krea [dev]. Notice the variety of face shapes, ethnicities and hair color.

FLUX.1 Krea [dev] has taken all the parts we like about FLUX.1, excellent image quality, prompt following, etc and increased the diversity of identities and styles.

But Does it Train?

Since FLUX.1 Krea [dev] can make wonderful portraits out of the box, the question becomes can it be personalized.

FLUX.1 [dev] is probably the most popular model to fine-tune for headshots, but can FLUX.1 Krea [dev] do better?

Here we try the same prompt on LoRA trained on FLUX.1 [dev] and Flux.1 Krea [dev].

Flux.1 [dev] left, Flux.1 Krea [dev] center, example training image right

On the right we have an image generated with a LoRA for FLUX.1 [dev] and in the center we have the Krea LoRA version. On the left is an example image from the dataset.

A few things to note. The center image looks more like the training images face shape, but it is has also learned the skin text and reflectance. The FLUX.1 [dev] LoRA, is too shiny. FLUX.1 [dev] has features that are not present in the original images (more well defined creases on the face, where the original is softer).

The FLUX.1 Krea [dev] LoRA is not only preserving the identity better, it is able to generate novel backgrounds and clothing. It has learned the face without overtraining to the point it can't produce diverse images, and notable advantage over FLUX.1 [dev].

Styles Anyone?

Not only is FLUX.1 Krea [dev] great for personalization, but it is also a excellent target for style training.

Below are examples of our Thomas Cole style dataset, from our earlier style training blog, trained for FLUX.1 [dev] and FLUX.1 Krea [dev].

Top row is the Flux.1 Krea [dev] lora, bottom row is the Flux.1 [dev] lora

Both Flux.1 [dev] and the Krea variant are able to learn the Thomas Cole style well. However, the Krea LoRA displays more diversity.

A Bright Future

Black Forest Labs newest release is a great addition to the FLUX.1 family. Out of the box, the model displays more diversity, which transfer to training people and styles.

We can't wait to see what our users come up with, and shoutout to friends at Black Forest Labs and Krea for the great model.