More video on OpenAlchemy: Wan 2.2, lip-sync, and H3 with your own audio

Ten days ago we opened video generation with MiniMax H3. Since then four more video models have gone live on the same /v1/videos API, all running on Grid compute, and H3 itself learned to take audio. Together they cover what a short film needs: a shot that moves, a character who speaks or sings in time, and a voice you chose.

The new models

Wan 2.2 speech-to-video (wan2.2-s2v) — a keyframe plus an audio track. The character lip-syncs to the audio and moves with it, for as long as the audio lasts. This is what drives singing music videos and characters reading their lines in their own voice.

Wan 2.2 image-to-video (wan2.2-i2v) — a keyframe plus a prompt, 2–10 seconds at 16 fps. Its motion is laid out without the speed-up LoRA for the first steps, so a subject told to run actually runs instead of drifting in slow motion.

Wan 2.2 Animate (wan2.2-animate) — a keyframe plus a driving video: the character in the keyframe performs the motion from the video.

LatentSync 1.6 (latentsync-1.6) — give it a finished clip and a new speech track, and the mouth is re-synced to the new words. Useful for dubbing a shot into another language.

MiniMax H3 now takes audio

A ref clip can be given a voice as <Audio 1>, and the character speaks its line in that voice. A t2v or i2v clip can be given an audio track as its soundtrack, anchored from the first frame. In image-to-video the keyframe stays the first frame, so the character, its clothes and the place are exactly the ones you drew.

# An image-to-video shot that says a line you recorded:
# the keyframe is the first frame, the audio is the soundtrack.
curl https://api.openalchemy.io/v1/videos \
  -H "Authorization: Bearer $OPENALCHEMY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax-h3","mode":"i2v","duration":5,
       "first_frame_url":"https://.../keyframe.jpg",
       "audio_url":"https://.../line.wav",
       "prompt":"One continuous take from the first frame. She looks up and says the line."}'

# Lip-sync from a keyframe and a voice track (Wan 2.2 S2V):
curl https://api.openalchemy.io/v1/videos \
  -H "Authorization: Bearer $OPENALCHEMY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"wan2.2-s2v","first_frame_url":"https://.../keyframe.jpg",
       "audio_url":"https://.../line.wav","prompt":"She sings, smiling."}'

As before, jobs are asynchronous: poll GET /v1/videos/{id} and download from /content. Billing is per output second, held while the shot renders and refunded if it fails. Try them without code in the Grid Playground → Video.

For GPU providers

Engine Desktop 0.6.25 ships all of these models, and its speech engine now recovers on its own if an instance stops answering. Existing installs update automatically; new providers can get it from the Download page.