Skip to content

Reference images and voices#

Reference image#

Every shoot has a reference image. What it does depends on the preset family:

  • Image-to-video (LTX): the reference is the first frame. The video starts from it and moves according to the prompt.
  • Reference-to-video (MiniMax H3): the reference is a face and appearance anchor. The model builds the shot from the prompt and keeps the person looking like the reference.

The aspect-ratio rule

The aspect ratio of the reference must match the output aspect ratio of the preset. A mismatched image is cropped silently before generation; the result then ignores parts of the prompt and drifts from the reference. Crop the image in the Edit tab first.

Picking from the gallery. The picker filters by board, searches the image metadata (prompt text and more) and keeps loading as you scroll, so an old image is never out of reach. A reference shoot's Add several from the gallery picks up to the free slots in one go; the numbers on the thumbnails are the order the references get.

Choosing a face reference. A sharp, front-lit, eye-level frame of the face at the size it should appear on screen. Avoid sunglasses, motion blur and crowded frames. Two face references per shoot is the tested combination. A shoot accepts up to nine pictures, and a location, a pose or a composition sketch can take the extra slots: in a five-picture test (two faces, a room plate, a kneeling mannequin as the pose guide and a storyboard sketch as the composition guide) the take followed the room, the pose and the framing of the sketch. Every picture adds render time, and more people with pictures give the model more faces to keep apart.

Voice references#

Speech presets (MiniMax H3) can anchor a voice. Record or import a short clean clip of the speaker in the Voice references panel of the shoot and attach it to the shoot. Two voice anchors per shoot are fine and cost only a few percent of render time.

  • The clip should be the voice alone: no music, no room echo, a few seconds of natural speech.
  • The line to be spoken still goes into the prompt in quotes; the reference only shapes the timbre.
  • Without a voice reference the model picks a voice per take, so it changes between takes.

Identity across shots#

Reference images give a likeness lottery: roughly three takes out of four keep the face. For a character that appears in many shots, train a character LoRA and build a character set for the scene; the LoRA carries the identity even in shots without a reference.

Reference shoots can chain clips: clip 1 is rendered from the references, each continuation clip from the last frame of the previous clip without them, so the voice references anchor clip 1 only — unless Keep the references in the continuation clips is ticked on the shoot, which carries the references and the voice references into every clip. See Continuity and chaining.