Skip to content

Character LoRAs#

A character LoRA teaches the video model what a person looks like, so their features hold across takes and shots, even without a reference image. Each character keeps a bank of trained LoRAs on their card.

Train one#

  1. On the character card, click Train character LoRA. Training Studio opens with a project for this character and a dataset built from the base reference and the angles.
  2. Review the dataset, add extra frames from the gallery if you have good ones, and start the run. See Train a character LoRA.
  3. When the run completes, the LoRA appears in the bank of the character with its trigger word, strength, date and the run it came from. The bank keeps the last ten.

Use one#

  • Video: add the LoRA to the default LoRA list of a scene and build the character set. The default strength is 0.5; strength 1.0 breaks motion and lighting.
  • Prompts: for video, mention the trigger word (chr_<name>) when describing the character, together with the description used in training. Image LoRAs (Krea 2, FLUX.2) are used without it; the bank shows them as no trigger in the prompt.
  • Images: character LoRAs for video are trained for the video model and do not load in Image Studio. For images, train an image LoRA instead; see Train an image LoRA.
  • Group pictures: an image LoRA gives its face to everyone in the frame. Generate the group without it, then give the character to one person (or a different character to each) with Apply a character to a person.
  • Overview: the LoRAs tab of the Models area lists every LoRA, image and video, with the character it belongs to; the All LoRAs link on the character card opens it.

The LoRAs tab: one list, filtered by image or video

What a LoRA does and does not fix#

It fixes the lottery: with a LoRA the face, hair and signature clothing stay the same from clip to clip. It does not fix a bad reference image, a prompt that contradicts the description, or a voice; voices need a voice reference.