Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) technical reference #
Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) is Google’s generally available (GA) multimodal model for conversational video generation, multimodal video editing, first-and-last-frame interpolation, video extension, and standalone video upscaling via the Interactions API.
1. Technical specifications #
| Property | Value |
|---|---|
| Canonical model ID | gemini-omni-1.2-flash (GA release date: November 5, 2026; retirement: November 5, 2027 or later) |
CLI aliases (scripts/omni.py) | gemini-omni-1.2-flash, omni-1.2-flash, omni-1.2, 1.2, flash-1.2 |
| API surface | Interactions API (client.interactions.create / v1beta1/projects/{PROJECT_ID}/locations/global/interactions); GenerateContent is not supported |
| Vertex AI / Enterprise region | global |
| Supported modalities | Text (input and output), Image (input only: JPEG, PNG, WebP, HEIC, HEIF), Video with synchronized speech, music, and sound effects (input and output: MP4, MOV as video/quicktime, WebM, MPEG, MPG, WMV, 3GP, FLV; standalone audio input/output is not supported) |
| Maximum input tokens | 131,072 tokens |
| Maximum output tokens | 65,536 tokens (260,640 tokens on Vertex AI / Gemini Enterprise Agent Platform) |
Output video duration (duration) | 3s to 15s (default 10s, whole seconds) for text-to-video, single-image-to-video, reference-to-video, and video editing; 3s to 10s for 2-image first-and-last-frame interpolation |
Video extension (extend) | 3s to 10s per turn (default 10s), using up to the final 10s of the input video as context (up to 30s total in Gemini API; up to 40s concatenated on Gemini Enterprise / Vertex AI) |
Standalone video upscaling (upscale) | Upscales 360p input videos to 720p or 1080p, and 720p input videos to 1080p or 4k (<= 15s at 24 FPS, total_input_tokens: 0) |
| Reference images | Up to 6 in Gemini API (10 max on Gemini Enterprise / Vertex AI; up to 3 images per character for up to 2 characters) across all durations (3s to 15s) |
| Reference videos | Up to 3 reference videos (up to 3s each) for 3s to 10s output clips; at most 1 reference video (up to 3s long) for 11s to 15s output clips |
| Input video for editing / upscaling | 1 input video up to 15s long (720p or 1080p recommended for editing; 360p or 720p for upscaling) |
| Resolutions and aspect ratios | 360p, 720p (default), 1080p, 4k at 24 FPS; 16:9 (default), 9:16 |
| Delivery modes | Inline base64 (delivery: "inline", default), Files API URI (delivery: "uri", recommended for > 4 MB / > 720p), or Cloud Storage (gcs_uri on Gemini Enterprise / Vertex AI) |
| Watermarking | Invisible SynthID watermark embedded in generated video frames and audio |
Deprecated control tags:
Do not include control strings such as <FIRST_FRAME>, <LAST_FRAME>, <IMAGE_REF_0>, or <VIDEO_REF_0> in text prompts for gemini-omni-1.2-flash. Describe how each image or video should be used in natural language instead, and optionally set --task (generation_config.video_config.task) if needed.
2. Gemini Enterprise / Vertex AI (global) and Gemini API authentication #
gemini-omni-1.2-flash runs on the global location in Gemini Enterprise / Vertex AI (https://aiplatform.googleapis.com/v1beta1/projects/{PROJECT_ID}/locations/global/interactions) as well as the Gemini Developer API (GEMINI_API_KEY). scripts/omni.py defaults --location to global when initializing ADC:
# Option A: Gemini Enterprise / Vertex AI (ADC)
gcloud auth application-default login --billing-project YOUR_PROJECT_ID
export GOOGLE_CLOUD_PROJECT="YOUR_PROJECT_ID"
export GOOGLE_CLOUD_LOCATION="global"
export GOOGLE_GENAI_USE_ENTERPRISE="true"
# Option B: Gemini Developer API
export GEMINI_API_KEY="your-api-key"
uv run scripts/omni.py \
--enterprise \
--project "$GOOGLE_CLOUD_PROJECT" \
--model "gemini-omni-1.2-flash" \
--resolution "1080p" \
--duration "15s" \
--prompt "A cinematic drone shot flying over a coastal lighthouse at sunset, crashing ocean waves and warm orchestral strings." \
--output "lighthouse_15s.mp4"3. Task modes (generation_config.video_config.task) #
By default, gemini-omni-1.2-flash infers the video task from your natural-language prompt and input media. Pass --task on the CLI or set generation_config.video_config.task when you want to pin the task (and always when running upscale):
| Task value | Description | Duration range | Required inputs |
|---|---|---|---|
text_to_video | Generate a new video with synchronized audio from a text prompt | 3s to 15s (default 10s) | Text prompt |
image_to_video | Animate 1 starting image (3s to 15s) or interpolate between 2 frames (3s to 10s) | 3s to 15s (1 image) / 3s to 10s (2 images) | 1 or 2 images + prompt |
reference_to_video | Generate a new video guided by character, object, style, or motion references | 3s to 15s | Up to 6-10 images and/or up to 3 videos (<=10s) or 1 video (11s-15s) + prompt |
edit | Edit an existing video while preserving unmentioned visual and audio elements | 3s to 15s | 1 input video (or --previous-interaction-id) + up to 2 ref images + prompt |
extend | Extend the end of an existing video using up to its final 10s as context | 3s to 10s per turn | 1 input video (or --previous-interaction-id) + up to 2 ref images + prompt |
upscale | Upscale a 360p video to 720p/1080p or a 720p video to 1080p/4k | Preserves input duration (<= 15s) | 1 input video (or --previous-interaction-id) + target --resolution |
4. Python SDK (google-genai) patterns for gemini-omni-1.2-flash #
The SDK examples below initialize genai.Client(enterprise=True, location="global", ...); omit those arguments to use GEMINI_API_KEY via genai.Client().
Text-to-video (15 seconds with synchronized dialogue and audio) #
import base64
from pathlib import Path
import google.auth
from google import genai
credentials, project_id = google.auth.default()
client = genai.Client(
enterprise=True,
project=project_id,
location="global",
credentials=credentials,
)
interaction = client.interactions.create(
model="gemini-omni-1.2-flash",
input=(
"[0-5s] Wide shot of an observatory dome opening under a starry desert sky with low mechanical hum. "
"[5-10s] Medium shot of an astronomer looking through a brass eyepiece. "
'[10-15s] She smiles and says in a hushed voice: "There it is, right on schedule."'
),
response_format={
"type": "video",
"resolution": "1080p",
"aspect_ratio": "16:9",
"duration": "15s",
},
generation_config={
"seed": 42,
},
)
if interaction.output_video and interaction.output_video.data:
Path("observatory_15s.mp4").write_bytes(
base64.b64decode(interaction.output_video.data)
)First-and-last-frame interpolation (3s to 10s) #
import base64
from pathlib import Path
from google import genai
client = genai.Client(enterprise=True, location="global")
first_b64 = base64.b64encode(Path("first_frame.jpg").read_bytes()).decode("utf-8")
last_b64 = base64.b64encode(Path("last_frame.jpg").read_bytes()).decode("utf-8")
interaction = client.interactions.create(
model="gemini-omni-1.2-flash",
input=[
{"type": "image", "data": first_b64, "mime_type": "image/jpeg"},
{"type": "image", "data": last_b64, "mime_type": "image/jpeg"},
{
"type": "text",
"text": (
"Use the first image as the starting frame and the second image as the ending frame. "
"Create a smooth time-lapse transition from daytime construction to an illuminated skyscraper at night."
),
},
],
response_format={
"type": "video",
"resolution": "1080p",
"aspect_ratio": "16:9",
"duration": "10s",
},
generation_config={
"video_config": {
"task": "image_to_video",
},
},
)
if interaction.output_video and interaction.output_video.data:
Path("interpolation.mp4").write_bytes(
base64.b64decode(interaction.output_video.data)
)Standalone video upscaling (task: "upscale") #
import base64
from pathlib import Path
from google import genai
client = genai.Client(enterprise=True, location="global")
source_b64 = base64.b64encode(Path("draft_720p.mp4").read_bytes()).decode("utf-8")
interaction = client.interactions.create(
model="gemini-omni-1.2-flash",
input=[
{"type": "video", "data": source_b64, "mime_type": "video/mp4"},
],
response_format={
"type": "video",
"resolution": "4k",
"delivery": "uri",
},
generation_config={
"video_config": {
"task": "upscale",
},
},
)5. Performance and latency optimization tips #
- Iterate at
360por720p, then upscale: Draft and refine camera movement, dialogue timing, and scene edits at360por720p, then run--task upscale --resolution 1080p(or4k) on the final clip. Upscaling bypasses text reasoning (total_input_tokens: 0) and preserves the exact motion and audio of your draft. - Use stateless unary mode (
--no-store) for single-turn renders: When you do not need follow-up conversational turns (--previous-interaction-id), pass--no-store(store=False, background=False, stream=False) to enable server-side unary optimization and avoid multi-hop continuation prefill overhead. - Use
delivery: "uri"or--gcs-urifor high-resolution clips: Videos above720por10soften exceed4 MB; pass--delivery urior--gcs-uri gs://your-bucket/out/to avoid large inline base64 payloads.
