Reference Images
Reference images are the foundation of a stable character. Influgen uses them to extract face embeddings, build face-lock guidance, and select stronger inputs for later workflows like LoRA training.
Primary vs. reference images
When you upload a character's first image, Influgen automatically treats it as the primary reference if no type is provided. Later uploads default to reference.
That distinction matters because the face-lock system starts from the first strong image it can rely on.
What happens on upload
When a reference image is uploaded, Influgen:
- stores the file in configured media storage
- creates a reference-image record for the character
- appends the new URL to the character profile
- extracts a new face embedding
- updates the character's stored face embedding
This means reference management is not cosmetic. Every upload can influence future generation behavior.
[screenshot: Reference image manager with primary photo pinned first and additional references in a sortable grid]
Supported upload field names
For direct API or form integrations, Influgen accepts multipart files under any of these field names:
fileimagereference_image
If you are using the web or mobile app, you do not need to think about field names. The UI handles them.
How many images should you upload?
For most characters, start with:
- 1 primary image
- 4 to 8 additional references
Add more only when they improve coverage. A bloated reference set with inconsistent styling often hurts more than it helps.
What makes a good reference image
- face takes up enough of the frame
- eyes are visible
- lighting is clean
- skin texture is not over-filtered
- age and styling match the intended persona
- the image feels like the same person you want to publish as
What to avoid
- sunglasses
- strong beauty filters
- extreme side profiles as your only inputs
- heavy makeup shifts that change bone structure perception
- mixed-age source sets
- inconsistent hair color between uploads
Managing the set over time
As the character matures, your reference set should improve, not just grow. Remove weak images, promote a stronger primary if needed, and use the best-performing generations as candidate material for LoRA training only after manual review.
The goal is not to create a giant archive. The goal is to maintain a tight, high-signal identity set.