Gemini Omni 1.1 Flash Reference-to-Video API by Google

Gemini Omni 1.1 Flash — Reference to Video

Model ID: google/gemini-omni-1.1-flash/reference-to-video

Gemini Omni 1.1 Flash is Google DeepMind's natively multimodal model for video generation and editing. This variant generates a new scene conditioned on reference media — up to 10 reference images and, new in 1.1, up to 3 short reference video clips — so a character, product, or art direction stays consistent across every generation.

Overview

Gemini Omni Flash is a single transformer that processes text, images, audio, and video simultaneously rather than stitching separate modality-specific pipelines together. The result is output that is more cohesive, more consistent across turns, and more controllable than a chained generate-then-dub workflow.

What sets it apart from earlier video models is that it natively generates audio with every video — dialogue, ambience, music, and sound design are produced together with the picture rather than added afterward. The model is grounded in Gemini's real-world knowledge, so it reasons about physics, narrative logic, culture, and visual composition to produce results that feel intentional and cinematic. Every generated clip carries an imperceptible SynthID watermark alongside C2PA Content Credentials for provenance verification.

What's new in 1.1

Version 1.1 is a control-and-fidelity upgrade over the first Omni Flash release:

• Flexible resolution output — new 4K and 1080p rendering for finished work, plus a fast 360p draft mode for previewing a shot before committing to a full-quality render.
• Longer continuous generation — shots can be extended segment by segment to a total of up to 40 seconds of coherent video.
• Precise shot start and end control — first and last frame can both be supplied, producing smoother camera moves, scene transitions, and seamless loops.
• Character and style consistency — a new video reference capability (clips of up to 3 seconds) substantially improves subject and art-direction stability. This is the endpoint that exposes it.

How the five endpoints relate

AtlasCloud exposes Gemini Omni 1.1 Flash through five model IDs — text-to-video, image-to-video, reference-to-video, video-edit, and video-extend. All five route to the same underlying gemini-omni-1.1-flash model and differ only by the input modality they accept, corresponding to the model's task parameter (text_to_video, image_to_video, reference_to_video, edit, extend). This endpoint maps to reference_to_video.

Inputs

This variant takes a text prompt plus reference media. At least one of reference_images or reference_videos must be supplied; both may be used together.

• Prompt — natural-language description of the new scene to build around the references, up to 20,000 characters.
• reference_images — 1 to 10 images used as character, scene, or style references. PNG, JPEG, JPG, or WebP, each up to 20 MB, supplied as public URLs or base64-encoded images.
• reference_videos — 1 to 3 MP4 clips, each up to 3 seconds long, used as motion, character, or style references. Each item must be a publicly accessible URL.

Unlike still references, a reference video carries motion as well as appearance — a walk cycle, a mannerism, a camera behaviour, or an animation style — into the new scene.

Referencing media from inside the prompt

Rather than describing an uploaded asset in words, you can point at it directly:

• <IMAGE_REF_N> — the Nth entry in reference_images, 0-based, so the first image is <IMAGE_REF_0>.
• <VIDEO_REF_N> — the Nth entry in reference_videos, 0-based, so the first clip is <VIDEO_REF_0>.

Bindings may also be declared up front with a [# References <TAG>@MediaN] block at the start of the prompt, where the @MediaN handles are 1-based (@Image1 is the same item as <IMAGE_REF_0>).

Example: The character in <IMAGE_REF_0> walks through the rain with the gait and camera energy of <VIDEO_REF_0>.

Key Capabilities

• Subject and style consistency — carry a referenced character, object, or look across scenes and generations.
• Multi-reference conditioning — blend up to 10 reference images and up to 3 reference clips to guide subject, scene, motion, and style at once.
• Motion references — a short clip transfers movement and animation feel, not just appearance.
• Inline reference tagging — address individual assets from the prompt with <IMAGE_REF_N> and <VIDEO_REF_N>.
• Native audio generation — every clip is rendered with a synchronized soundtrack (speech, music, effects) driven by your description.
• Draft-to-final resolution ladder — iterate cheaply at 360p, then re-render the shot you like at 720p, 1080p, or 4K.
• Adjustable reasoning — the thinking_level control trades latency for quality on complex prompts.
• Reproducible results — set a fixed seed to reproduce or iterate on a specific generation.

Input Parameters

Conditional requirement: at least one of reference_images or reference_videos must be present. A single generation call produces at most 10 seconds; to reach the model's 40-second ceiling, chain the result into the video-extend endpoint.

Prompting Tips

• Give the model clean references: an unobstructed, well-lit view of the subject beats a busy frame.
• Use several images of the same character from different angles to lock identity more firmly than a single frame can.
• Reserve reference_videos for what stills cannot express — a specific gait, gesture, animation style, or camera behaviour.
• Tag references inline instead of re-describing them; <IMAGE_REF_0> is less ambiguous than "the woman in the red coat" when several subjects are in play.
• Fix a seed once you have a look you like, so subsequent episodes in a series stay aligned.

Use Cases

• Consistent characters — keep the same protagonist, mascot, or presenter across a series of clips.
• Brand identity — reproduce a product, logo, or visual style consistently across marketing videos.
• Style transfer — apply the look and feel of reference art, or the motion feel of a reference clip, to a newly described scene.
• Episodic content — maintain visual continuity across many generations in a storyline.
• Personalized media — generate videos featuring specific subjects supplied as references.

Pricing

Billing combines a per-second output charge that depends on the requested resolution with a small per-reference ingestion charge for each supplied image and video clip.

SKUs

Formula

cost = max(3, duration) × resolution_rate
     + count(reference_images) × $0.00168
     + count(reference_videos) × $0.0261

Where resolution_rate is $0.041 for 360p, $0.11 for 720p, $0.16 for 1080p, and $0.31 for 4k.

Notes and examples

• The output charge is per second of generated video, with a 3-second minimum — a shorter request is still billed as 3 seconds.
• Reference charges are per item, not per second: each image adds $0.00168 and each clip adds $0.0261 regardless of the clip's length.
• duration defaults to 10 and resolution defaults to 720p when not supplied; both reference arrays default to empty.
• A 10-second 720p video with 3 reference images and no reference video costs 10 × $0.11 + 3 × $0.00168 = $1.10504.
• A 10-second 720p video with 3 reference images and 1 reference clip costs 10 × $0.11 + 3 × $0.00168 + 1 × $0.0261 = $1.13114.
• A 10-second 360p draft with 2 reference images costs 10 × $0.041 + 2 × $0.00168 = $0.41336.
• A 10-second 4K video with 5 reference images and 3 reference clips costs 10 × $0.31 + 5 × $0.00168 + 3 × $0.0261 = $3.1867.