Skip to main content

Gemini Omni Flash (Video)

The GeminiOmniAdapter wraps Google's Gemini Omni Flash preview model (gemini-omni-flash-preview) for fast multimodal video generation and conversational video editing via the Interactions API.

Overview​

Unlike Veo (which uses models.generate_videos), Omni uses client.interactions.create(...). It accepts text, images (and in the broader Omni family, audio/video inputs) and returns short clips (about 3–10s at 720p / 24 FPS) with natively generated audio. Prior generations can be refined with natural-language edit turns using previous_interaction_id.

References:

Authentication​

Uses the same key as Nano Banana / Veo:

export GEMINI_API_KEY="your_api_key"

Quick Start​

from dotenv import load_dotenv
load_dotenv()

from tryon.api.omni import GeminiOmniAdapter

adapter = GeminiOmniAdapter()

# Text-to-video
video = adapter.generate_text_to_video(
    prompt="A fashion model walking a runway under soft studio lights",
    aspect_ratio="9:16",
)
with open("outputs/omni_t2v.mp4", "wb") as f:
    f.write(video)

# Image-to-video
video = adapter.generate_image_to_video(
    image="data/model-1.jpg",
    prompt="The model begins walking toward the camera, soft cinematic motion",
    aspect_ratio="16:9",
)

# Conversational edit (reuse the prior interaction id)
edited = adapter.edit_video(
    prompt="Dim the lights and add a slow dolly-in",
    previous_interaction_id=adapter.last_interaction_id,
)
with open("outputs/omni_edit.mp4", "wb") as f:
    f.write(edited)

API Reference​

Class: GeminiOmniAdapter​

Methods​

  • generate_text_to_video(prompt, aspect_ratio="16:9", previous_interaction_id=None) → bytes
  • generate_image_to_video(image, prompt, aspect_ratio="16:9", reference_images=None, previous_interaction_id=None) → bytes
  • edit_video(prompt, previous_interaction_id=None, aspect_ratio=None) → bytes

aspect_ratio supports 16:9 and 9:16. After each successful call, adapter.last_interaction_id is updated so you can chain edits without re-uploading the prior clip.

When previous_interaction_id is passed to generate_text_to_video / generate_image_to_video, the request is treated as an edit task.

opentryon CLI / MCP Server​

# Text-to-video
opentryon video-generate --model gemini-omni \
  --prompt "A fashion model walking a runway" --aspect-ratio 9:16

# Image-to-video (passing --image switches the method)
opentryon video-generate --model gemini-omni \
  --prompt "Animate a slow walk toward camera" --image photo.jpg

# Conversational edit
opentryon video-generate --model gemini-omni \
  --prompt "Dim the lights" --previous-interaction-id <id>

MCP tool: video_generate_gemini_omni.

See Also​