Skip to main content

Custom LoRA Training

Custom LoRA is the upgrade path for characters that need stronger identity retention than the default IP-Adapter tier can provide.

When to train a LoRA​

You probably need a LoRA when:

  • the character will publish at high volume
  • you want tighter facial resemblance across many scenes
  • the niche depends on beauty, fashion, or recognizable close-ups
  • you plan to use motion-heavy workflows like talking head or motion transfer

If the default face-lock is already good enough, do not train one just because it exists. LoRA is a power tool, not a requirement.

Plan requirement​

LoRA training is available on Pro and Agency plans.

Training requirements​

The current training rules are:

  • minimum images: 10
  • maximum images: 30
  • training steps: 1000 to 1500
  • default training steps: 1200
  • rank range: 16 to 32
  • default rank: 16
  • default base model: fal-ai/flux-lora
  • default training cost: 500 credits
  • face-score threshold for eligible images: 0.85

If training images do not meet the similarity threshold, the run is blocked before credits are wasted.

[screenshot: LoRA training page showing uploaded training images, face scores, and readiness state]

1. Start with good references​

Choose images that are:

  • sharp
  • front or three-quarter view
  • evenly lit
  • consistent in age and overall look
  • varied in outfit and background, but not in identity

2. Upload training images​

Influgen accepts multipart uploads for training images and stores them as a separate training set. If a run is already in progress, new uploads are blocked until it finishes.

3. Review face scores​

Each candidate image is scored against the character identity. Images below the threshold should be removed or replaced.

4. Start training​

You can let Influgen use defaults or specify:

  • base_model
  • training_steps
  • lora_rank
  • a selected subset of image_ids

5. Wait for readiness​

Training moves through states such as queued, training, ready, or failed. Once ready, the character's active face-consistency tier becomes lora.

What changes after training​

Once a model is ready, Influgen stores:

  • the active model URL
  • the trigger word
  • the training config
  • the training image count
  • the charged credit timestamp

The generation pipeline then prefers LoRA-backed generation with face-lock fallback instead of reference-only guidance.

Common reasons training fails​

  • fewer than 10 valid training images
  • more than 30 images uploaded
  • a training run is already in progress
  • storage for training assets is unavailable
  • insufficient credits
  • low face scores across the candidate set

Should every character get a LoRA?​

No. A good rule:

  • one-off or experimental characters: stay on IP-Adapter
  • high-performing or revenue-bearing characters: consider LoRA
  • premium brand work or large-scale publishing: LoRA is usually worth it

The best moment to train is after the character has already proven itself, not before.