MiniMax H3 Max (Fal)

OpenTryOn’s first third-party hoster adapter. Fal jointly released MiniMax H3 Max with MiniMax and hosts a post-trained stack with T2V, I2V, reference-to-video, and a dedicated lip-sync/dubbing endpoint.

This is not the MiniMax first-party V2 API. Official MiniMax H3 Max (--model minimax-h3-max, MINIMAX_API_KEY) is T2V / first-last I2V only. Use Fal when you want R2V, lip-sync, or Fal’s queue.

CLI model	Host	Modes	Key
minimax-h3-max	MiniMax V2	T2V, I2V	MINIMAX_API_KEY
fal-h3-max	Fal queue	T2V, I2V, R2V	FAL_KEY
fal-h3-max-lipsync	Fal queue	Image + audio → lip-synced video	FAL_KEY
minimax-h3	MiniMax V2	T2V, I2V, R2V (Python), 2K	MINIMAX_API_KEY

No open weights for H3 Max. Local H3-Base is --model minimax-h3-local.

Docs: T2V, I2V, R2V, Lip Sync, queue.

Environment

export FAL_KEY=...          # https://fal.ai/dashboard/keys
# export FAL_API_KEY=...    # alias
# export FAL_QUEUE_BASE_URL=https://queue.fal.run

Auth header is Authorization: Key {FAL_KEY} (not Bearer).

CLI

opentryon video-generate --model fal-h3-max \
  --prompt "A fashion model walking a runway at dusk, camera tracking" \
  --duration 5 --resolution 768P --ratio 16:9

opentryon video-generate --model fal-h3-max \
  --image look.jpg --prompt "Gentle fabric motion as the model turns" \
  --duration 6

opentryon video-generate --model fal-h3-max \
  --prompt "Walk toward camera, stop on the mark" \
  --last-frame end.jpg

opentryon video-generate --model fal-h3-max \
  --prompt "Image 1 is the model. Keep her identity while she walks the runway." \
  --reference-image look.jpg \
  --duration 5 --resolution 768P

# Lip-sync / dubbing: portrait + audio in, matched mouth movement out (no prompt)
opentryon video-generate --model fal-h3-max-lipsync \
  --image portrait.jpg --audio line.wav \
  --resolution 1080P --enable-transcription

Flag	Notes
--duration	Integer 5–15 seconds (default 5)
--resolution	480P or 768P (default 768P)
--ratio	T2V: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 (not adaptive). R2V may use adaptive. I2V follows the keyframe.
--image	First frame (switches to I2V)
--last-frame	Last frame (alone or with --image)
--reference-image / --reference-video / --reference-audio	R2V lists. At most 12 files total. Audio cannot be the only reference. Mutually exclusive with --image.
--prompt-expansion	balanced (default) or quality
--no-safety-checker	Disable Fal’s safety checker
--seed	Optional

fal-h3-max-lipsync flags

No --prompt — only an image and an audio track. Resolution ceiling is higher than T2V/I2V/R2V (up to 2K).

Flag	Notes
--image	Portrait to animate (aspect ratio must be 0.4–2.5) — required
--audio	Track to sync mouth movement to (≥5s; clipped to ~14.8s) — required
--resolution	480P, 768P, 1080P, or 2K (default 768P)
--enable-transcription	Transcribe the audio to guide lip-sync accuracy
--no-safety-checker	Disable Fal’s safety checker
--seed	Optional

Python

from tryon.api.fal import FalH3MaxAdapter

adapter = FalH3MaxAdapter()
open("t2v.mp4", "wb").write(
    adapter.generate_text_to_video(
        prompt="A fashion model walking a runway at dusk",
        duration=5,
        resolution="768P",
        ratio="16:9",
    )
)
open("i2v.mp4", "wb").write(
    adapter.generate_image_to_video(
        image="look.jpg",
        prompt="Gentle turn toward the camera",
        last_frame="end.jpg",
    )
)
open("r2v.mp4", "wb").write(
    adapter.generate_text_to_video(
        prompt="Image 1 is the model. Keep her identity while she walks.",
        reference_image=["look.jpg"],
    )
)
open("lipsync.mp4", "wb").write(
    adapter.generate_lip_sync(
        image="portrait.jpg",
        audio="line.wav",
        resolution="1080P",
    )
)

Notes

• Jobs go through Fal’s queue (POST https://queue.fal.run/minimax/h3-max/... → poll status_url → download video.url).
• MCP tools: video_generate_fal_h3_max and video_generate_fal_h3_max_lipsync (generated from the registry).
• Planner: fal h3 max / fal-h3-max pin T2V/I2V/R2V. h3 max lip sync / fal-h3-max-lipsync pin the dubbing endpoint. A bare h3 max still pins first-party minimax-h3-max.