MiniMax H3 Talking Photo API

One portrait plus one voice clip in, a talking video out — driven by the MiniMax H3 multimodal engine. The person in the photo speaks the audio with frame-accurate lip movement, while identity, outfit, framing and background stay pixel-faithful to the source image: only the face animates, nothing drifts. 480p or 720p, at fixed lengths of 5, 10 or 15 seconds, priced per clip: 480p from $0.07 and 720p from $0.13, known before you submit.

Pricing (per clip)

Resolution	5s	10s	15s
480p (default)	$0.07	$0.11	$0.22
720p	$0.13	$0.22	$0.44

Fixed per-clip pricing — not metered per second, and independent of the actual audio length. Charged only on success, refunded on failure. At 16:9 the 480p tier measures 864×480 and the 720p tier 1376×768 (about 1 megapixel).

Versus AI Lip-Sync (ai-lipsync, $0.02 per audio second): for very short audio (under ~3s) ai-lipsync is cheaper; for longer audio or 720p output this model wins, and the price is known before you submit.

Authentication

All requests carry the API key in the header:

Authorization: Bearer YOUR_API_KEY

Generate a talking video

POST/api/v1/video/generations·GET/api/v1/video/generations?task_id=

# Talking photo: 1 portrait + 1 driving audio -> talking video
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3-talking-photo",
    "image": "https://example.com/portrait.jpg",
    "audio_url": "https://example.com/voice.mp3",
    "duration": 5,
    "resolution": "480p",
    "ratio": "16:9"
  }'

# Poll until completed (result files are kept 7 days)
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"

# The audio must FIT the chosen duration tier (5 / 10 / 15s) — we probe it at
# create time and return a clear 400 if it does not, before anything is billed.
# Shorter audio is fine: the person finishes speaking and idles naturally.

Parameters

Field	Required	Type	Description
model	Yes	string	minimax-h3-talking-photo
image	Yes	string	The portrait, exactly 1 (public URL or base64). Real photos, illustrations and AI-generated faces all work. images[0] / image_urls[0] are also accepted.
audio_url	Yes	string	The driving audio, exactly 1 (public URL or base64, mp3 / wav). audio_urls[0] / audio_list[0] are also accepted.
duration	Yes	number	5 / 10 or 15 — pick the tier that covers your audio: we probe the audio length at create time and return 400 if it does not fit. The field name seconds is also accepted.
resolution	No	string	480p (default) or 720p.
ratio	No	string	16:9 (default) / 9:16 / 1:1 / 4:3 / 3:4 / 3:2 / 2:3 / 21:9. aspect_ratio is also accepted.
prompt	No	string	Optional, guides the delivery; a neutral description is used when omitted.
callback_url	No	string	We POST the result to this URL on completion; omit it and poll the GET endpoint instead.

Notes

Async: create returns a taskId; poll GET /v1/video/generations?task_id= until completed, or pass callback_url. Measured turnaround: roughly 2-4 minutes for 480p·5s, up to ~13 minutes for 720p·15s. Result files are deleted after 7 days — download or re-host promptly.
Audio shorter than the tier is fine: the person finishes speaking and settles into a natural idle (blinking, micro-movement), the track pads with silence, and nothing loops. Audio longer than the tier is a 400 at create time — no task, no charge.
Parameter errors (duration outside 5/10/15, resolution or ratio outside the enums, more than one image or audio) return a synchronous 400 at create time.
Prompts and outputs go through content moderation; a rejection is treated as a failure and refunded. Use only portraits you have the rights/consent to depict.
For a variant whose output length follows the audio exactly, billed per second, see AI Lip-Sync.

Try it in the Playground

FAQ

How much does MiniMax H3 Talking Photo cost?

Fixed per-clip tiers on apimodels.app: 480p at $0.07 / $0.11 / $0.22 and 720p at $0.13 / $0.22 / $0.44 for 5 / 10 / 15 seconds. The price is independent of the audio length and known before you submit. Only successful generations are charged; failures are refunded.

How do I call the MiniMax H3 Talking Photo API?

POST /api/v1/video/generations with model minimax-h3-talking-photo, exactly one image (the portrait), one audio_url (the voice, mp3/wav) and duration 5, 10 or 15. Optional: resolution (480p default or 720p), ratio (16:9 default, 8 values) and a prompt. Poll GET /api/v1/video/generations?task_id=xxx or pass callback_url. One API key covers every model on apimodels.app.

What happens when the audio is shorter or longer than the clip?

Shorter is fine: the person finishes speaking and settles into a natural idle — blinking and micro-movement, no looping — while the audio track pads with silence. Longer is rejected at create time: we probe the audio duration and return a clear 400 before anything is billed, so pick the 5/10/15s tier that covers your audio.

How is this different from AI Lip-Sync?

AI Lip-Sync bills $0.02 per audio second and the output length follows the audio exactly. Talking Photo uses the MiniMax H3 engine with fixed 5/10/15s tiers and per-clip pricing, plus a 720p option and 8 aspect ratios. Very short audio (under ~3s) is cheaper on AI Lip-Sync; longer audio or 720p output favours Talking Photo.