Make a Talking Video from One Picture and a Voice
LTX-2.5 Audio-to-Video animates a single image so the person speaks along to the voice you give it — lips, face, and gestures follow the audio, and your audio stays exactly as it was. Made for presenter clips: a photo on a green background plus a voiceover or TTS line becomes a 24 fps talking video, for voices up to 30 seconds. Runs on CNAPS infrastructure; no external API key required.
What it does
LTX-2.5 Audio-to-Video (A2Vid) from Lightricks turns one image and a voice into a talking video. The voice drives the motion: mouth shapes, head movement, and hand gestures follow it, and the original audio is muxed into the result untouched.
By default it picks a vertical (768×1344), landscape (1344×68), or square (1024×1024) canvas to match your image, or you can choose one. It renders at 24 fps, for voices from 0.5 to 30 seconds. Generation takes a few minutes. Put the subject on a flat green background to composite them later with Chroma Compose.
To generate a new scene with its own sound from a prompt, use LTX-2.5. To re-sync the lips of an existing talking video, use LatentSync 1.6.
Problem it solves
- Presenter clips for ads — One photo and a TTS voiceover become a spokesperson video
- Talking heads for explainers — Lip-synced product explainers without a shoot
- Other languages from the same photo — Swap the voice track to localize the same presenter
- Green-screen presenters — Lay the presenter over screen recordings or product shots
- Audio stays exact — Your voiceover is muxed in as-is, so timing and quality never change
Input/Output
Input
- Image (required) — the first frame: the person who will speak. A flat green background composites best
- Sound (required) — the voice to speak, 0.5–30 s; kept as-is in the result
- Text (optional) — a description of the motion and scene
Parameters (main view)
Option | Values · what it does |
Output size | auto (default) / 768×1344 (vertical) / 1344×768 (landscape) / 1024×1024 (square). Auto picks the canvas closest to the image's aspect ratio. With fit set to crop, a mismatched canvas cuts the subject; use pad to avoid that |
Advanced Options
Option | Range / default · what it does |
Stage-1 denoising steps | 10–60, default 30 (the checkpoint default). Stage 2 always uses the distilled schedule. Cost is roughly linear — about 5 s per step |
Random seed | Default -1 (random each run). Set a fixed value to reproduce a result |
Warm-up before the first word | 0–2 s, default 0.8. The model keeps roughly the first 0.5–0.8 s close to the still image, so a voice that starts immediately gets almost no lip motion on its first word. The warm-up lets the face start moving before the voice, then is trimmed away — output timing and length are unchanged. 0 disables it. Voices near the 30 s cap get a shorter warm-up |
Fit (crop / pad) | crop (default): the image is centre-cropped to the canvas, so a picture with a different aspect loses its edges. pad: the image is grown to the canvas aspect with its border colour — extra height on top only, extra width split between the sides — so nothing is cut. Meant for stills on a flat backdrop such as chroma green; on a busy photo the padding shows as a plain band |
Output
- Video — the talking video at 24 fps with the voice muxed in
How to Use
- Connect an Image Loader with a front-facing photo of the presenter — ideally on a flat green background.
- Connect the voice: a Sound Loader, or a TTS node such as Text to Speech (Qwen3 Voice Design). Keep it under 30 seconds.
- (Optional) Connect a Text Input describing the motion and scene (e.g., "friendly presenter, natural hand gestures").
- Leave Output size on auto, connect Video Viewer, and run. Expect a few minutes per clip.
- Composite the green-screen presenter with Chroma Compose, then add captions and finish the mix.
Example Workflow
Flow: Image Loader + Sound Loader (+ optional Text Input) → LTX-2.5 Audio-to-Video → Video Viewer.
Generated voice: Text Input → Text to Speech (Qwen3 Voice Design) → LTX-2.5 Audio-to-Video sound port.
See the full pipeline in the Talking Presenter Ad recipe.
Use Case Tips
- Portrait-cut presenters — If the photo is cut off at the bottom edge, use pad: extra height is added on top only, so the presenter stays anchored to the bottom.
- Lips on the first word — Keep the warm-up at 0.8 s; raise it toward 2 s if the first word still looks stiff.
- Speed vs. quality — Lower the stage-1 steps for drafts (~5 s saved per step), then return to 30 for the final.
- Longer scripts — Split narration over 30 s into segments, generate each, and join them with Video Concat.
How it compares
What you need | Use |
A person in one photo speaking a voice you provide | LTX-2.5 Audio-to-Video |
A new shot with generated sound, from text or a first frame | LTX-2.5 |
An existing talking video re-synced to a new voice | LatentSync 1.6 |
Video with native stereo audio from reference images and clips | MiniMax-H3 Ref2VA |
Model Source
- CNAPS model page: LTX-2.5 Audio-to-Video
- Hugging Face: Lightricks/LTX-2.5
- GitHub: Lightricks/LTX-2
Compliance & Provenance
Item | Detail |
Provider | Lightricks (open weights) · Specialized |
License | LTX-2.x Community License |
EU AI Act risk class | Limited Risk |
Art. 50 transparency | Required — AI-generated outputs are marked. A talking video of a real person's likeness may also need disclosure as manipulated content |
For platform classification, acceptable use, and output marking, see the CNAPS AI Policy.
Run it now
Try it in the Talking Presenter Ad recipe, or open CNAPS Studio, find LTX-2.5 Audio-to-Video in the Component panel, and start with the example workflow above.