Gemini Omni Flash 1.1
What is Gemini Omni Flash 1.1?
Gemini Omni Flash 1.1 is Google's multimodal video model, released August 27, 2026 as the replacement for the May Gemini Omni Flash. Text, images, video, and audio all sit in the model context, and the output is high-resolution video with audio. Google orders the name as Gemini Omni 1.1 Flash; on Unifically it runs as google/gemini-omni-flash-1.1-video and google/gemini-omni-flash-1.1-video-edit.
The headline change is memory. The May model referenced only the final second of prior video when building on earlier context. 1.1 reads up to 10 seconds, so characters, lighting, and scene layout hold together across follow-up edits instead of drifting after the first change.
Everything else that made the first Gemini Omni Flash useful carries over. Media is context: a product still, a phone video, a character reference, a voice preset, and a prompt can all steer one result, and the next instruction builds on what came before. The model also keeps Gemini's real-world knowledge, which matters for scenes with cause and effect: gravity, fluid motion, object interaction, and short explainers where the video has to show an idea instead of only matching a style.
The 1.1 endpoints accept the same request body as the May endpoints. Migrating is a one-line change: swap the model string.
What's new in Gemini Omni Flash 1.1
10-second context window
The model reads up to 10 seconds of prior video context, up from one second in the May release. Edits and reference-led generations keep character identity, lighting, and scene layout consistent across changes.
#1 on Arena text-to-video
As of August 28, 2026, one day after release, gemini-omni-1.1-flash holds #1 on the Arena text-to-video board at 1515 Elo, above the May model at 1512 and FLUX 3 Video at 1495. On image-to-video it holds #2 at 1488 Elo, three places above the May model.
First-and-last-frame keyframes
Give the model a start image and an end image and it generates the video between them: camera moves, transitions, and loops. Pass start_image_url plus end_image_url in frame mode.
Half-price 360p drafts
A native 360p resolution renders drafts at half the 720p price. Test prompts at 360p, then rerun the one you want to deliver at 720p or with in-task upscale to 1080p or 4k.
Drop-in migration
Same request body as the May endpoints: prompt, references, characters, voice, aspect_ratio, duration, resolution, seed. Change the model string to google/gemini-omni-flash-1.1-video and existing requests keep working; end_image_url and 360p are new options on top.
Best for
Mixed-context video prompts
Use it when the prompt needs to combine written direction with images, source video, characters, and voice choices.
Reference-led short videos
Use image and character references when the result needs to follow a subject, product, outfit, or setting from uploaded assets.
Existing video edits
Edit mode is built for one source video plus a prompt: restyle it, change the scene, or replace a subject with a character reference.
Multi-pass edit loops
The 10-second context window is built for follow-up edits. Each new instruction builds on the last result without losing the character or the scene.
Concept explainers
Use it when a short video needs to show a process, scientific idea, or cause-and-effect scene with grounded motion.
Voice-guided character videos
Use a request-level voice preset for generation, or per-character voice fields in edit mode when the video needs speech direction tied to a subject.
Controlled reruns
Seed support helps rerun a prompt and reference setup when you need a close repeat for review or testing.
Use cases
Turn a rough product video into a cleaner ad concept by giving the source video, brand stills, and a short prompt, then asking for lighting and background changes. Generate a character scene by uploading one to three character references and calling them out with @Character1 in the prompt. Rework a creator video for a new setting: keep the subject, change the environment, and ask for a camera move that fits the format. Build a compact explainer, such as a clay-style process video or a chain reaction with clear physical motion. Test a storyboard by combining sketches, a character reference, a voice preset, and written scene direction, then rerun with the same seed once the direction is close.
Limitations
Google's 1.1 release also introduced scene extension up to a cumulative 40 seconds and short video references in generation. Neither is on the Unifically endpoints yet. Generation here runs prompt, frame, and reference modes at 4 to 10 seconds; editing works on one uploaded source video.
Audio input stays narrow. Voice presets and per-character voice fields are the supported audio path; the model does not analyze arbitrary uploaded audio, and voice editing of existing speech is not supported.
360p and 720p generate natively; 1080p and 4k output is upscaled inside the same task, not generated natively at those sizes. A 360p video can only be upscaled to 720p later.
Speech edits, captions, brand marks, and people-centric changes still need a review pass. Generated videos carry a SynthID watermark, but the watermark does not replace human QA.
Gemini Omni Flash 1.1 vs Gemini Omni Flash
On the Arena text-to-video board, 1.1 took #1 at 1515 Elo one day after release, edging out the May model at 1512. The bigger gap is on image-to-video: 1.1 holds #2 at 1488 Elo while the May model sits at #5 with 1463, so the quality gain shows most when an input image has to carry through the whole video. The video-edit board does not list 1.1 yet; the May model held #4 there. All numbers pulled August 28, 2026.
The practical difference for API users is the retirement: the May model IDs no longer accept new tasks, and the 1.1 endpoints take the same request body. There is no reason to stay behind.
When to use Gemini Omni Flash 1.1
Use Gemini Omni Flash 1.1 when a video job needs more than a prompt: source video, still references, character references, first-and-last-frame control, voice direction, grounded scene logic, or a targeted edit of one uploaded video. Use Veo 3.1 when the job depends on Extend or Upscale as separate calls.
What Gemini Omni Flash 1.1 can do
#1 on the Arena text-to-video board
One day after release, Gemini Omni Flash 1.1 holds #1 on the Arena text-to-video board at 1515 Elo from 1,762 votes, above the May model at 1512 and FLUX 3 Video at 1495. On image-to-video it holds #2 at 1488 Elo. Numbers as of August 28, 2026.
Speech and audio generated with the video
One start image and one prompt produce the finished result: the man moves naturally, delivers his line out loud, and the cafe ambience sits behind him. Dialogue is written straight into the prompt, no separate audio model.
Edit an existing video with one instruction
The source is an oil-painting style video with a bottle on the table. One sentence swaps the bottle for an apple while the brushwork, lighting, and motion of the rest of the scene stay put. This is the edit loop the model is built around.
Build a scene from reference images
Two stills in different painterly styles go in as references with a one-line prompt, and the model composes a coherent cinematic scene that draws on both. References steer subject and style without locking the first frame.
Direction-following text-to-video
A shot-by-shot fashion editorial brief: five named poses, an 85mm slow dolly, specified lighting and color grade. The model holds the studio, the subject, and the camera plan across the full 10 seconds.
API examples
Call Gemini Omni Flash 1.1 from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/video/google/gemini-omni-flash-video.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "google/gemini-omni-flash-1.1-video",
"input": {
"prompt": "Make @Character1 and @Image1 dance in a neon studio with energetic camera movement.",
"reference_image_urls": [
"REFERENCE_IMAGE_URL"
],
"reference_characters": [
{
"image_urls": [
"CHARACTER_IMAGE_URL"
],
"name": "Dancer"
}
],
"duration": 10,
"aspect_ratio": "9:16"
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished result.
FAQs
People also ask
Gemini Omni Flash 1.1 is Google's multimodal video model, released August 27, 2026. Text, images, video, and audio can all be part of the model context, and the output is high-resolution video with audio. It replaces the May 2026 Gemini Omni Flash model.
The main change is context. The May model referenced only the final second of prior video when building on earlier context; 1.1 reads up to 10 seconds, so characters, lighting, and scene layout stay consistent across edits. It also adds first-and-last-frame keyframes and a half-price native 360p draft resolution. One day after release it took the
Two. Use google/gemini-omni-flash-1.1-video for new text-to-video, frame-mode (a start image, optionally with an end image), or reference-to-video generations. Use google/gemini-omni-flash-1.1-video-edit when you already have one source video and want to change it with a prompt. The older google/gemini-omni-flash-video and google/gemini-omni-flash-video-edit IDs are retired; swap the model string and keep the same request body.
Yes. Output is video with audio. Generation mode supports a request-level Google voice preset when at least one image or character reference is attached, and edit mode supports per-character voice fields, including tuned voices with performance direction.
Yes. The edit model takes exactly one source video up to 30 seconds and 1GB, plus an edit prompt. You can ask for scene changes, background swaps, camera moves, action changes, and style changes, and use start_frame and end_frame to pick the exact source range to work from.
The generate model supports 4, 6, 8, or 10 second videos in 16:9 or 9:16. The edit model accepts one uploaded source video up to 30 seconds and returns an edited result based on the selected frame range.
As of August 28, 2026, one day after release, it holds
Yes. Google uses SynthID on generated images and videos, so output carries a provenance watermark. It does not replace human review for people-centric or speech-heavy edits.
