Upload one photo with a visible face and one audio track you recorded or own. The tool builds a lip-sync MP4 from that pair. It is not a still-plus-audio overlay and not generic image-to-video motion.
Audio · required
A clearly visible face in the still gives the lip-sync a better chance. This is guidance, not a hard gate. Sign in before uploading.
Use a still with a visible face and a clean uploaded track. The generated MP4 replaces this preview when it is ready.
An AI talking photo is a short MP4 in which a still photograph appears to speak in time with a voice you already have. You upload one photo and one audio file. The generator estimates mouth motion from that track and writes new video frames. The still is not simply played with the audio laid on top. Generic image-to-video motion, which invents camera travel or weather from a prompt, is a different job.
The useful input is a photograph where a face is clearly visible, well lit, and not cropped through the mouth. Sunglasses, heavy occlusion, extreme profile, or a tiny face in a crowd make lip-sync harder. That is a quality hint, not a hard gate. The workbench will still accept a file that meets size and type rules. Results stay probabilistic. Teeth, tongue, and the edges of the lips can drift. Identity around the jaw can soften. A second person in the frame may pick up leftover motion. Inspect the download before you publish it.
Banana Pro AI charges from the audio duration the server measures during upload, not from the seconds a browser player reports. 32 credits per second at 720 and 48 credits per second at 1080. Billed length is one second through 60 minus one. Tracks at 60 seconds or longer are rejected. A variable-bitrate MP3 with no readable duration metadata can fail even when the player preview looks fine. Optional hint text may run to 300 characters. v1 does not add text-to-speech, voice clone, a seed, or a mask.
These frames are source stills and still posters for a talking photo brief. They are not generated lip-sync clips and not live recordings. Use them to judge whether your own still is a similar kind of input.
A catalog or landing still of a person holding a product, recorded later as a spoken pitch. The talking photo path is for a face that already exists in the photograph plus a track you recorded in a quiet room. Keep the product label readable in the still. Do not expect the generator to invent a second camera angle or a new pair of hands.
A formal portrait used with a short greeting you recorded yourself. A talking photo can carry a remembered voice onto a still you have the right to use. It cannot restore a person who is not in the photograph, and it cannot promise a perfect match to every family memory. Keep the upload lawful. Confirm rights before you spend credits.
A presenter still for a course, FAQ, or support clip. The talking photo output should keep the same clothing and background while the mouth follows your uploaded track. If you need a new room, a walking shot, or a cut between slides, use image-to-video or a filmed session instead of this workbench.
Prepare the still and the track first. Credits are quoted only after the server measures the audio. Generation is a tracked task you can reopen from My Creations.
Choose a JPG, PNG, or WebP up to 10 MB, with each side at least 256 pixels. A face that fills a reasonable part of the frame, looking toward the camera, is the honest starting point for a talking photo. Crop before you upload if the subject is tiny.
Upload MP3, WAV, AAC, OGG, or M4A up to 10 MB. Prefer a single speaker and a readable file. The server returns billed seconds. The on-page player is a preview only. A warning appears when measured length is over 15 seconds. Hard rejection is 60 seconds or longer.
Default output is 1080 at 48 credits per billed second. 720 costs 32 credits per billed second. Confirm that you own the photo and the audio, or that you have permission to use them. The generator will not add a second voice or clone a voice from a prompt.
Sign in, spend the quoted credits, and wait. Status checks run every few seconds and pause after fifteen minutes. The same request id can resume without a second charge. If the task fails, reserved credits return. Retry from My Creations restores only the hint and resolution. You must upload the photo and audio again.
Watch the whole clip. Check the first and last phonemes, teeth, and whether a second face moved. Download through My Creations. Treat the file as a draft until you have watched it. A talking photo is not a substitute for a filmed interview when legal or archival accuracy matters.
Pick the path that matches the assets you already have. A talking photo needs one still and one uploaded track. A talking avatar is a persistent character. Image-to-video is motion from a still. Traditional filming captures a real performance on a timeline.
| Talking Photo vs Talking Avatar vs Image-to-Video vs Traditional filming | Best for | Control | Tradeoff |
|---|---|---|---|
| Talking photo | A real photograph of one person plus a recorded track you can upload | You choose the still, the audio file, and 720 or 1080. Duration comes from the track | Lip-sync can drift. No TTS, no voice clone, no extra camera move beyond what the model invents around the mouth |
| Talking avatar | A reusable digital presenter that should appear across many scripts | Character setup and later takes share an identity you maintain over time | You are not locking output to one existing photograph. Avatar pages are a different product |
| Image-to-video | A still that should gain camera motion, weather, or a slow product orbit from a written prompt | Start frame, optional end frame, duration, and a motion prompt | The mouth is not driven by your audio. Do not use it when the job is a talking photo |
| Traditional filming | Interviews, legal statements, and any take where the real performance must exist on a camera card | Lighting, lenses, retakes, and an editor on a local timeline | Highest fidelity and highest production cost. No model will replace a shoot you still need |
The output is a generated lip-sync MP4. It is not a recovered recording of the day the still was taken. It is not a lossless overlay of audio onto the JPEG. If the face is small, turned, or covered, sync quality usually drops. If the audio is noisy, clipped, or full of overlapping speakers, the mouth motion will show it. Credits are billed from server-measured audio seconds. A browser preview duration is not the bill.
v1 requires a user-uploaded audio file. There is no text-to-speech box, no voice clone from a sample, and no duration slider on generate. Tracks at 60 seconds or longer are rejected. Length above 15 seconds is allowed under that cap and may look less stable. Variable-bitrate MP3 files without duration metadata may be rejected even when they play in the preview.
Limits and storage
A signed-in run creates a tracked task. You can reopen the result from My Creations after a dropped connection without another charge for that request id. The current storage backend has no automatic expiry for uploaded or generated objects. Delete in My Creations soft-deletes the task record from account history. It does not delete the underlying object or a provider copy. Review the Privacy Policy and the Terms of Service before you upload sensitive faces or private speech.
My Creations · Privacy Policy · Terms of Service
U.S. Copyright Office, Copyright and Artificial IntelligencePublic materials on copyright questions involving artificial intelligence. Not legal advice.Stay on this talking photo page when you have one still and one uploaded track. Use the related tools when the job is motion from a still, a broader video generator, or an avatar identity.
Animate a still with a motion prompt when the mouth should not follow an audio file
Image to Video AIBrowse the wider video generator when you are still choosing a starting asset
AI Video GeneratorBuild a reusable avatar when you need a character rather than one photograph
AI Avatar GeneratorStraight answers about audio-driven lip-sync, credits, length limits, and what this talking photo page will not do.
A lip-sync MP4 from one still photo and one uploaded audio track. The mouth is driven by that track. The still is not played as a slideshow with sound on top.
Yes in v1. The talking photo workbench requires a photo and a user-uploaded MP3, WAV, AAC, OGG, or M4A file. There is no TTS field and no voice clone.
From billed seconds the server measures on audio upload. 32 credits per second at 720. 48 credits per second at 1080. Default resolution is 1080. The player duration is a preview only.
Shorter than 60 seconds. A warning appears over 15 seconds. 60 seconds or longer is rejected. Files above 10 MB are rejected. A VBR MP3 without readable duration metadata may fail.
A visible face is strongly recommended for a talking photo. It is not a hard software gate. Occluded or tiny faces usually produce weaker lip-sync. Inspect the result.
No. Image-to-video invents motion from a prompt. A talking photo locks the performance to your uploaded audio. Use image-to-video when you want camera travel without a voice track.
Upload a photo you may use, upload the audio, pick 720 or 1080, and inspect the lip-sync MP4 before you share it.
Open the workbench