Skip to content
Bud's Wiki

Gemini Omni 1.1 Flash lets you build with more control

TL;DR

Section titled “TL;DR”

Google’s follow-up release to Gemini Omni (Introducing Gemini Omni) — a “1.1 Flash” version aimed at developer/production use via the Gemini API and Agent Platform. Adds four production-oriented controls on top of the launch model: scene extension of an existing clip in 10s increments up to 40s total, first- and last-frame keyframe specification for controlled transitions, resolution tiers spanning 360p (drafting) through 4K (final), and video-reference conditioning (up to 3s per reference) alongside text and image inputs. No architectural detail, parameter count, or benchmark numbers are disclosed; this is a product blog, not a technical report.

Key claims

Section titled “Key claims”
  • Scene extension now analyzes up to 10 seconds of prior clip context — the launch model referenced only the final second — and can extend a video in 10s increments up to a cumulative 40s total [§Extend scenes for longer storytelling].
  • First-frame + last-frame keyframe conditioning generates continuous video between two specified keyframes, positioned as the recipe for smooth transitions, camera orbits, zooms, and seamless looping clips [§Specify first and last frames].
  • A 360p tier is offered as a draft/preview resolution at ~60% faster wall-clock generation and one-third the cost of 720p, with 1080p and 4K available for final output [§Draft videos more efficiently in 360p, §Upscale up to 4K resolution].
  • Video-reference conditioning accepts up to three video references (up to 3s each) in the multimodal input, used to transplant motion or character identity from the references into a generated scene under a text prompt [§Add video references in your multimodal input].
  • API surface exposes these as parameters on a single gemini-omni-1.1-flash model behind client.interactions.create, with previous_interaction_id chaining used for scene extension and response_format.resolution selecting the tier [code snippet, §Extend scenes for longer storytelling].
  • Distribution at launch is broad: Gemini API in Google AI Studio, Gemini Enterprise Agent Platform API for enterprises, plus Google Flow for all AI Plus/Pro/Ultra subscribers and scene extension in the Gemini app [§Build with Gemini Omni 1.1 Flash Today].

Method

Section titled “Method”

Not disclosed. The post is a product/API announcement; no architecture, tokenizer, training recipe, evaluation numbers, or watermarking details are included beyond what was published for the launch model card (Introducing Gemini Omni). Public code snippets show the API pattern only. The four capabilities (extension, keyframes, resolution, video references) are all exposed as request-time parameters rather than as separate model heads, consistent with the launch framing of Omni as a single natively-multimodal model rather than a pipeline of specialists — but nothing in the post confirms whether these are truly a single-model surface or a routed set of specialized sub-models.

Results

Section titled “Results”

None reported. No head-to-head numbers against Veo 3, Sora 2, Wan 2.7, MiniMax H3, or the launch Omni Flash. No T2V/I2V/first-last-frame benchmark tables. Qualitative demos only — a drummer/saxophonist/ballet dancer whip-pan (first-last-frame), microscopic diatoms (360p draft), fish/chipmunk/maple-leaves (4K), and a three-character dance replacement (video references). Performance claims are limited to the two comparatives already noted: ~60% faster and 3× cheaper for 360p vs. 720p, and the 10s→40s extension increase [§Draft videos more efficiently in 360p, §Extend scenes for longer storytelling].

Why it’s interesting

Section titled “Why it’s interesting”

This is the first production-side iteration of Gemini Omni after its May 2026 launch (Introducing Gemini Omni), and it lands as a feature update rather than an architecture bump — Google is choosing to compete on controllability (keyframes, references, extension, resolution tiers) rather than raw fidelity or duration, which is a different Pareto position from Veo 3 (Veo 3 Tech Report) and MiniMax H3 (MiniMax H3 — Open-Weights Omni-Reference Video-Audio Generation Model Card). The first-frame/last-frame control primitive is exactly what CineTechBench: A Benchmark for Cinematographic Technique Understanding and Generation identified as the substantially-more-informative variant of camera-motion conditioning for I2V models — CineTechBench found that first+last-frame conditioning materially reduced translation error over first-frame-only, so this is Google shipping the operating point that recent benchmarks have argued for. The scene-extension mechanic — 10s of prior context, extendable in 10s chunks to 40s — sits on the same context-length/quality-tradeoff Pareto (Context Length / Quality Trade-off in Video Generation) as the KV-partitioning (Context Forcing: Consistent Autoregressive Video Generation with Long Context) and learned-memory-compression (PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference, Pretraining Frame Preservation in Autoregressive Video Memory Compression) recipes, but exposed as a chunked API rather than a rolled-out streaming generator like LongLive: Real-time Interactive Long Video Generation. The 40s ceiling is short of what open long-video systems (LongVie 2 at 5 min, PackForcing at 2 min) claim — this is either a product-side cap or an actual quality limit, and the post doesn’t say.

See also

Section titled “See also”