Introducing the FLUX Portrait Trainer

Team Features & Labels

• 3 min read
Introducing the FLUX Portrait Trainer

Today we are happy to release a new trainer designed specifically for high quality portraits: the FLUX Portrait Trainer. The Fal FLUX portrait trainer creates striking portraits, with fine details, bright highlights in the eyes, better representation of different face sizes, better prompt following while maintaining high resemblance.

Improved Fine Details.

The first thing you’ll notice with the FLUX Portrait trainer is the images seem to “pop” off the screen. A big reason for this is preservation of FLUX’s ability to generate fine details, like single hairs, sharp highlights in the eyes and more realistic skin texture.

Two images generated with the same model using the FLUX Portrait trainer. Notice the sharp eyebrows and highlights in the eyes.

Better Prompt Following

The FLUX Portrait trainer is able to follow more complex prompts with higher fidelity. This allows for more accurate inventive scenes, but also increases relighting capabilities.

In the example below, the dataset only included scenes that were shot during the day, making it difficult for the FLUX general trainer to relight the scene for to “pitch black” as the prompt required, but the FLUX Portrait trained model had no issue.

Two generations with the same seed and parameters. FLUX Portrait trained model is on the left and the general trained model is on the right. The prompt is “kptz woman in a pitch black room. I single shaft of volumetric light slants across her right eye.” Dataset provided by CheeseAI, used with permission.

The prompt following extends to modifications to the identity of the subject. Here we prompt “kptz old woman” and we see that only the FLUX Portrait trained model is able to adhere to the prompt.

Two generations with the same seed and parameters. Left was using the FLUX Portrait trained model and is able to accurately follow the prompt "kptz old woman." The General trained model on the right produces an image consistent with the training data, but ignores the prompt.

Improved Small Faces

Smaller faces in generations tend to lose resemblance. We’ve added a multi-resolution training option which helps to maintain identity even as the faces get smaller.

“kptz man riding a shark” is the prompt. General trained model on the left and Portrait trained model on the right.

Also notice that shark is better represented in the scene with the FLUX Portrait trained model. In general the FLUX Portrait trainer causes less broken anatomy in scenes.

High Resemblance

All these features are great, but they are nothing without being able to capture high resemblance. Luckily the FLUX Portrait trainer maintains high resemblance even with dramatic scene and lighting changes.

More Trainers More Options

We think the FLUX Portrait Trainer is a great complement to our current general FLUX Trainer. When training people, the general trainer is faster and able to preserve the identity more predictably than the Portrait Trainer, however the image quality of Portrait is higher and it can generate more verastile prompts.

We think they are both great and depending on your particular application one might work better than other.

We are constantly researching how to improve fine-tuning. Even as we release this trainer, we’ve been testing new techniques to push the boundaries of what is possible for image generation. Checkout the FLUX Portrait Trainer here and keep an eye out for more fruits of our training research.