Omni Flash 1.1: Extending AI Video Past 10 Seconds
40 s is a chain of 10 s segments; 4K is an upscale of 720p; keyframes and video references on Google Flow's updated generation rail.
OMNI LabsUpdated
Key takeaways
A single Omni Flash 1.1 take caps at 10 seconds; reaching 40 seconds requires chaining sequential segments.
Extension works strictly at clip ends; it cannot insert backwards or add new dialogue to speaking subjects.
First-and-last frame control completed an 8 s descent at 1920x1080 on 2 September 2026, matching Veo 3.1 Fast.
Flow 360p draft mode costs half the credits of 720p, providing an economical path to validate camera pacing.
4K output is an upscale of an internal 720p render; fine detail not present at 720p is synthesized by the upscaler.
Scene extension mechanics and sequential chaining constraints
A single Omni Flash 1.1 generation on Google Flow is still capped at 10 seconds. A 40-second scene is a chain of 10-second segments, not one take, so a brief that asks for one unbroken 40-second take cannot be met on this rail.
Marketing claims surrounding generative video extensions frequently mislead production directors into expecting continuous, unbroken takes. Within Google Flow, Omni Flash 1.1 enforces a hard architectural limit of 10 seconds per individual generation. When a commercial brief calls for an extended 40-second sequence, the timeline must be assembled as a chain of sequential 10-second clips. Attempting to generate a continuous 40-second camera move in a single pass is technically impossible on this rail.
Extension only works at the end of a clip. It cannot insert in the middle or extend backwards, a clip with a speaking subject cannot take new dialogue, and an uploaded clip must be 10 seconds or shorter to be extended.
Operational boundaries govern how footage extensions execute. The Flow extension tool operates strictly unidirectionally from the final frame forward. It cannot extend footage backwards to reveal prior action, nor can it insert new narrative beats between existing cuts. Crucially, if an existing clip features a speaking subject, the extension engine cannot generate new dialogue. External reference media submitted for extension cannot exceed 10 seconds in duration.
First-and-last frame interpolation and the Veo signature
First-and-last-frame control worked on the live Flow panel on 2 September 2026. An orbit-to-driveway pair rendered as one continuous 8-second descent with no cut, and the identity held into the final frame.
Boundary keyframing provides essential structural discipline for complex spatial transitions. On 2 September 2026, our team tested Omni Flash 1.1 on a demanding camera transition connecting an orbital space plate to a vehicle parked in a luxury driveway. The model generated an unbroken 8 s descent, transitioning between keyframes without cuts or geometric warping while maintaining vehicle identity through to the final frame.
That same first-and-last-frame job came back at 8 seconds and 1920x1080, which is the signature of Veo 3.1 Fast. We do not know whether the panel routed it to Omni Flash, so the credit cost on the Omni tier is unread.
Analyzing backend output metadata revealed critical architectural ambiguities. Although Omni Flash was selected, the rendered file returned at exactly 8 seconds in 1920x1080—the unmistakable technical signature of Veo 3.1 Fast. Until telemetry confirms whether Flow routed the task to Veo or processed it natively on Omni, the true credit cost on the Omni tier remains unverified.
Our cloud-video file still lists Omni Flash as start-frame only. Our Omni 1.1 file says that rule may be dead. Until the live credit quote is read with both frame slots filled, we do not route keyframe jobs to Omni Flash.
Start-frame efficiency, selectable durations, and 360p drafting
A start-frame-only 4-second Omni Flash job held the vehicle's geometry, grille and emblem cleanly on 2 September 2026. Start-frame-only at 4 seconds is the known cheap path.
When camera transitions do not require boundary locking, start-frame conditioning represents the most cost-effective production route. In our 2 September 2026 trials, generating a 4-second take from a single initial frame preserved intricate vehicle grilles and hood emblems perfectly, establishing a dependable baseline for short commercial shots.
Google Flow Omni Flash 1.1 operational parameters and feature specifications · measured 2 Sept 2026
Operational Feature
Capability Specification
Technical Boundary
Studio Protocol
Scene Extension
Chain up to 40 seconds cumulative
Single generation capped at 10 seconds
Extend only at clip end; no backward extension
Keyframe Interpolation
Start and end frame support
Output returned at 8 s and 1920x1080 (Veo signature)
Read live credit quote before routing jobs
Resolution Scaling
4K output tier
Native render is 720p; 4K is an upscale
Finish 720p master with dedicated upscalers
Draft Mode Efficiency
360p preview rendering
Costs half the credits of standard 720p
Use 360p for rapid framing and camera approval
Reference Ingestion
Up to 10 images (IMAGE_REF_0 to 9)
Up to 3 video refs of 3 seconds each
Reference audio ignored; voice uneditable
Live panel verification completed 2 September 2026; 10-second duration unique to Omni rail.
Omni Flash 1.1 offers 4, 6, 8 and 10 second durations. It is the only model on the Flow rail that offers 10 seconds.
Flow's 360p draft costs half the credits of 720p. We had first written a third, which was wrong: the Flow model page says half.
Draft modes provide critical economic protection when testing complex motion. Flow's 360p draft setting cuts credit consumption in half compared to 720p. Using 360p allows directors to iterate camera velocity and subject blocking cheaply before committing credits to final renders.
Reference-guided prompting syntax and 4K upscaling realities
Reference-guided mode takes up to 10 image references, labelled in the prompt as IMAGE_REF_0 to IMAGE_REF_9, and up to 3 video references of 3 seconds each. The 3-reference limit we wrote earlier belongs to Veo 3.1 Ingredients, not to this mode. Audio inside reference videos is ignored and voice cannot be edited.
Conditioning complex scenes on external assets requires disciplined syntactic labeling. Omni Flash 1.1 accepts up to 10 distinct reference stills, addressed positionally within prompts as IMAGE_REF_0 through IMAGE_REF_9, alongside up to 3 short video references. Video reference audio is discarded by the model, and synthetic vocal tracks cannot be manipulated through prompt instructions.
4K on Omni 1.1 is an upscale. The model generates at 720p internally, so detail that is not in the 720p render is invented by the upscaler.
Understanding internal raster limits prevents unrealistic client expectations. Omni Flash generates video at a native internal resolution of 720p. Selecting 4K output does not engage native high-density diffusion; instead, the platform routes the 720p render through an upscaling filter. Fine fabric textures or typography absent in the 720p baseline cannot be recovered.
Prompt discipline: enforcing continuity and managing chain drift
For a first-and-last-frame prompt we write 'one continuous shot' and 'no scene cuts'. Without them Omni tended to build several shots. Omni has no negative-prompt field, and excluded terms were seen appearing in the output, so we write the positive state instead.
Preventing unscripted editorial cuts requires affirmative phrasing. When keyframing in Omni, omitting continuous shot instructions frequently causes the model to construct rapid internal cuts. Because Omni Flash lacks a negative prompt field, prompts must enforce continuity affirmatively by stating 'one continuous shot' and 'no scene cuts'.
Drift over a chain is not measured. Google published no consistency figures across a 40-second chain, so we assume drift accumulates with every segment until a test says otherwise.
In extended 40-second chained sequences, latent drift poses a constant threat. Because Google has published no empirical consistency metrics across multi-segment chains, our studio assumes that minor stylistic and anatomical variations compound with each 10-second link, requiring vigilant visual auditing.
