tendre.AI keeps a character's identity from one image to the next, and into video. What carries that identity depends on what the character has: a trained LoRA, the most faithful, or a text description.

A trained LoRA: the character's real identity

In My characters, Create a character opens a single profile; training is its LoRA training section. Give it 10 to 15 photos, real or generated, with a sharp face and above all variety: different places, outfits and lighting. The Front, Profile and Full body slots guide you, and preparing the photos has its own guide. If photos are missing, Complete makes the missing images from yours.

Then train it where you want:

On our servers: about 55 minutes. The price is shown on the button before you start, and credits are only taken if training succeeds.
On your graphics card, free: about 5 hours on an 8 GB card, much less on a bigger one. It needs at least 8 GB of video memory, about 32 GB of RAM and the training workshop, installed from Components (about 25 GB). Training saves a checkpoint every 250 steps, so it can resume after an interruption.

Once trained, connect the character in Create an image and pick its Character LoRA tile: it puts @ and the character's name in your scene, and that @ sends its trigger word (the character's name followed by "ai", editable on its profile as LoRA keyword), its age and its gender. Click the @ to see exactly what is sent. If the trigger word is wrong, the LoRA loads but stays silent and the render shows a stranger.

The Strength slider of its card sets how tightly the face is held: 1 by default, 1.5 at most. Higher locks the face, lower gives the scene more freedom.

Without a LoRA: the description

A character without a LoRA carries its identity in the Characteristics section of its profile: type, age, hair, eyes, build, distinctive features. The description goes in front of every scene. This keeps the described traits, not an exact face: for the same face every time, train a LoRA.

If your scene asks for something else, another hair color or closed eyes for example, the scene wins over the description.

Write the prompt around the character

Describe the scene, outfit, pose and lighting as usual: the trigger word or the description takes care of the identity. Keep the face big enough in the frame. In a wide full-body shot the face is small and the LoRA has less to hold on to: frame closer or raise the strength.

For two characters in one image, see Character duo.

Retouching

The character's LoRA also works when you retouch an image by instruction. It matters most when the edit has to redraw the face, to turn the head or change the expression: the LoRA lets the model rebuild the face from an angle the photo does not show.

The character sheet

Seven reference views of the character, one sheet per outfit or attitude. They carry its likeness in Create a video, in Ultra, and in Quality when the character does not play through its video LoRA; Create an image takes one of its images. See Character sheets.

In video

Three ways to keep the character in a clip:

Animate an image. Generate an image of the character with its LoRA, then animate it in Video: the starting image carries the identity. An image LoRA cannot be used directly in a video.
Train a video LoRA (LTX 2.5, online only). It holds the face and the voice for the whole clip, from text as well as from an image. The wizard takes the character's photos, a voice and 36 short clips made from the photos, or your own clips. The price is shown before training, and credits are only taken if it succeeds.
Create a video puts several characters, objects and a location in the same clip: four subjects and a location in Quality, up to 30 references in Ultra. They play from their character sheets (in Quality, one character can play through its video LoRA): a character without a sheet is rendered from a single image and looks much less like itself. See Create a video.

Pick the character's voice in the Voice section of its profile: it is preselected in Video for clips with sound.

Bring an existing LoRA

Import a LoRA, on the character's profile in My characters, takes a .safetensors file. The local engine must be installed, since the file goes into its folder. Only Krea 2 LoRAs are accepted. A LoRA made for another engine (SDXL, FLUX...) is refused: no engine in tendre.AI can load it.

When you generate online with a character whose LoRA is only on your computer, the app offers to send it online when you click Generate. Nothing is sent without your agreement.

With the Pro licence

Characters trained online, or sent online, are also available to the Pro API and to connected agents (Claude, ChatGPT and others): the agent lists them and uses each one with its trigger word. See Pro API and MCP (agents).

When a character drifts

Check the trigger word on the character's profile (LoRA keyword).
Raise the character strength, or show the face bigger in the frame.
Look at the training photos: a setting or an outfit present on every photo ends up in the identity. Vary them and retrain.

Keep a character consistent