Face Reference

Face Reference is a dedicated feature that maintains identity across Product to Model, Model Swap, Create Model, and Edit. It keeps generations aligned with a single face image and works more reliably and easily than methods like fine-tuning or training on a set of images.

Why Face Reference instead of fine-tuning

Earlier versions of FASHN supported consistency through fine-tuning (custom training) on about 8 to 12 images of an individual. Although training can work, it is difficult to do well. Each source image can introduce bias, and the underlying model becomes less flexible after fine-tuning. Customers also found it challenging to gather a complete and consistent training set in the first place, since they needed consistent results before they could produce the required images.

After several months of testing with customers, we found that anchoring identity to a single face reference produces more reliable results, avoids dataset collection issues, and maintains the flexibility of the base model.

How Face Reference works

Face Reference influences supported tools in two ways:

• It guides the base generation to match the reference identity, including traits such as ethnicity, inferred body shape, age, and other coarse attributes.
• It reinforces the likeness with special focus on facial features.

Face Reference adapts identity to new angles, expressions, hairstyles, and lighting. It keeps the identity consistent while still allowing the model to generate natural, flexible, and creative variations.

Where to use it

• Product to Model: Generates a model that reflects the face reference across the full body. Traits such as age and ethnicity from the reference flow through the entire output.
• Model Swap: Replaces the existing model with the face reference and updates the full body to match, including age shifts or ethnicity changes.
• Create Model: Builds a new model from scratch that resembles the face reference, with full-body alignment from the start.
• Edit → Face Swap: Includes a Face Swap variant focused on restoring identity rather than altering it. Use it when pose, background, or other edits distort the face. It can also swap identities, but the underlying model must already match the new face in skin tone and age.

Additional controls

Face Reference adds a control that adjusts how strongly the result follows the reference image versus the base image or prompt:

• All supported tools show Match Reference.
• You will also see Match Base or Match Prompt, depending on whether the input starts from an image or a prompt.

Example (Model Swap): If the current model faces left and smiles, but the face reference looks straight ahead with a neutral expression, the control lets you decide whether to keep the existing pose and expression or follow the reference more closely.

Example (Product to Model): The control decides how much the prompt affects the result. For instance, a prompt such as "a model looking up" can shift pose and expression unless you set the control to follow the reference image more strongly.

Preparing reference photos

• Even though the feature can adapt to new angles or expressions, results improve when the reference photo already resembles the expected output. In these cases, keep the Match Reference control at its default setting to maximize resemblance.
• Hair, makeup, and accessories should be close to the desired final look.
• Avoid heavy occlusions such as sunglasses, masks, or hands covering the face.

Helper tools

• Create Face: Designed to work with Face Reference. It generates realistic faces and consistent variations, such as different directions and expressions. These results appear in the quick select menu when you open the Face Reference picker, making it easy to choose a prepared face for a scene.
• Upscale: Use the 2x Upscale option in the Gallery to increase facial detail in your preferred reference images. Added realism carries over to Face Reference generations.

Effect on runtime and cost

Face Reference is computationally heavy. Generation times typically increase from about 10 seconds to around 40 seconds. Since compute time drives credit usage, enabling Face Reference increases the cost of a run from 1 credit to 4 credits.

Limitations

Face Reference is currently supported only with the Precise 1K resolution in the app and returns 1K outputs. Support for 2K and later 4K is planned.

Workflow summary