How Do You Make a Photo Talk?

What the model actually reads

InfiniteTalk analyses the waveform of your audio to predict phonemes, then moves the mouth to match. That order matters: the face follows the sound, not the other way around. The practical consequence is that a clean voice track with one speaker produces a convincing result, while a phone recording of a room with music in the background produces a face that chews at nothing.

The portrait supplies identity, not performance. Hairstyle, wardrobe, lighting and background are held steady across the take, which is why a plain, evenly lit headshot looks more convincing than a dramatic studio shot with a hard shadow cutting across the jaw.

Step 1: prepare the voice track first

Record or export the audio as a mono mix with as little reverb as you can manage. Room echo is the single most common reason a talking photo looks uncanny: the model is told to speak, and the room answers a fraction of a second later.

• One speaker, no overlapping voices. A conversation with two people will be mouthed by one face.
• No background music. Strip it, or record the voice again — no model separates a voice from a track well enough for lip sync.
• Consistent level. A track that starts quiet and ends loud makes the face look bored and then shouty.
• Supported formats are MP3, WAV, OGG and M4A, up to 40 MB.

Step 2: choose a portrait the model can lock onto

Frontal or three-quarter photos work best. The face needs to be large enough that the mouth is more than a few pixels wide, and nothing should cross in front of it — a microphone, a hand, a strand of hair. If the photo has more than one person, upload a mask so the model knows which subject should articulate; otherwise it will split its attention and animate the wrong face.

Step 3: pick resolution and length, then render

Runs are billed in 5-second blocks. Use 480p to test whether the voice track and portrait pair well; 720p costs roughly twice as much per block, so it belongs on the final delivery, not on the first attempt. Maximum clip length is 10 minutes, and a file longer than 600 seconds is rejected before you are billed.

The prompt field takes direction rather than dialogue: "gentle smile", "studio key light", "subtle nodding pacing". Keep it short — the words are a delivery note for the performance, not a script. The script is the audio.

What it costs

A 5-second 480p run is 180 credits. Because the gift at signup is 300 credits, a new account can produce one test clip and still have credits left for a photo animation run. Paid plans start at $4.90 a month for 300 credits, and one-time packs start at $9.90 for 500 — the audio work is yours either way, and re-recording it is free.

When the result still looks wrong

• Mouth moves but the words do not land: the audio has echo or music. Re-record dry.
• Face drifts between identities: the source photo is a profile or heavily retouched. Use a straighter, less processed shot.
• Jaw looks mechanical on long takes: shorten the clip and re-render the rest separately, rather than pushing one long take.
• Expression is flat: add a delivery note to the prompt and re-run the shortest segment first.