Video from Reference

Generate videos with consistent subject appearance using reference images via Google DeepMind's Veo 3.1 model on fal.ai.

See references/fal-api.md for setup, Python patterns, and error handling.

API Endpoint

fal-ai/veo3.1/reference-to-video

Parameters

Required

• prompt (string): Text description of the video scene and action
• image_urls (array of strings): URLs of reference images for consistent subject appearance (use fal_client.upload_file() for local files)

Optional

CLI Script

python3 scripts/fal_generate.py \
    --endpoint video-from-reference \
    --prompt "The character walks through a sunny park" \
    --images ref-front.jpg ref-side.jpg \
    --video-resolution 1080p \
    --output video.mp4

Tips

• Provide multiple reference images from different angles for better subject consistency
• Use clear, well-lit reference images with the subject prominently visible
• Describe both the scene environment and the subject's actions in the prompt
• More reference images generally lead to better subject consistency (2-4 recommended)
• Works best with distinct, recognizable subjects (people, pets, products)
• Upload local files with fal_client.upload_file() before passing to image_urls