Fal AI Media

Fal.ai Media Generation

Fal.ai provides GPU-backed AI inference for generative media. This skill covers the Sync Lipsync (sync-3) image-to-video and video-to-video pipelines, and the general fal.ai upload + run pattern.

Authentication

# Store in ~/.hermes/.env
export FAL_KEY="<your-api-key>"

# Format: a 36-char UUID, colon, then a 32-hex-char secret
# Example: 5136e6b5-aa91-4b0f-b77c-5ba08944bc48:0d30a9ba4385b238d52f9185deac29e7

Install the Python client:

pip install fal-client

Upload Files to Fal Storage

Fal has built-in file storage. Upload before running any model:

from fal_client import upload_file

image_url = upload_file("/path/to/image.jpg")
audio_url = upload_file("/path/to/audio.mp3")

Files get a https://v3b.fal.media/files/... URL that is accessible to fal models for ~24 hours.

Sync-3 Image-to-Video (Photo + Audio → Talking Head)

Best model for turning a static photo with a face into a talking video with lip-sync.

Model ID: fal-ai/sync-lipsync/v3/image-to-video

Required inputs:

NOT face_image_url (that's for the video-to-video variant). Use image_url.

Python (synchronous):

from fal_client import run
import os

os.environ['FAL_KEY'] = '***'

result = run(
    "fal-ai/sync-lipsync/v3/image-to-video",
    arguments={
        "image_url": "https://...",
        "audio_url": "https://..."
    },
    timeout=300  # 71s audio can take 5-10 min
)

video_url = result.get('video', {}).get('url', '')

Python (with progress updates):

from fal_client import subscribe
import time

start = time.time()

def on_queue_update(update):
    status = type(update).__name__
    pos = getattr(update, 'position', '')
    elapsed = time.time() - start
    print(f"[{elapsed:.0f}s] {status}" + (f" — positie {pos}" if pos else ""))

result = subscribe(
    "fal-ai/sync-lipsync/v3/image-to-video",
    arguments={
        "image_url": img_url,
        "audio_url": audio_url
    },
   
)

Via curl:

curl -X POST https://fal.run/fal-ai/sync-lipsync/v3/image-to-video \
  -H "Authorization: Key $FAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "image_url": "https://...",
    "audio_url": "https://..."
  }'

Sync-3 Video-to-Video (Existing Video + New Audio → Re-sync)

When you already have a video but want to re-sync the lips to different audio.

Model ID: fal-ai/sync-lipsync/v3 (or fal-ai/sync-lipsync for v2)

Required inputs:

Response Shape

The response is a dict. The video URL is typically nested:

video_url = result.get('video', {}).get('url', '')
# Or sometimes:
video_url = result.get('url', '') or result.get('video_url', '')

Processing Times

Long audio can time out synchronous calls. For audio >30s, use subscribe() with a progress callback or run asynchronously.

Pitfalls

• image_url vs face_image_url — The image-to-video variant uses image_url. Using face_image_url (the video-to-video param name) returns a 422 validation error.
• Timeout on long audio — The default CLI timeout (180s) is too short for 60s+ audio. Set timeout=600 or use subscribe().
• The photo needs a visible face — If the image has no clear face or is heavily stylized, the model may hang in InProgress indefinitely or produce a garbled result.
• fal.ai website is Next.js — The fal.ai/storage URL is a web page, not an upload API. Use the Python client's upload_file() instead.
• Result URL is nested — Don't assume result['url']. Check result['video']['url'] first.
• 403 on first run — You need credits on your account. Check credits at https://fal.ai/dashboard.

Related Models on Fal