Gemini Omni Flash Turns Prompted Briefs Into Full-Context Video Edits

Alec Wilcock presents Gemini Omni Flash inside ElevenCreative as a prompt-driven video editor rather than a filter. The operational pattern across the demonstrations is consistent: put a clip in the video reference slot, describe what should change, and explicitly describe what must remain fixed — the subject, movement, camera action, composition, scale, timing, face, voice, or audio, depending on the job.

That distinction is clearest in the restyling workflow. A claymation prompt is not just “make this claymation.” The shown prompt asks the model to restyle the video in claymation, change the color grading, clothes, and environment to match that style, and keep the main character, movement, camera action, composition, and scale exactly the same. Wilcock’s point is that the style instruction and the preservation instruction do different work: one changes the medium; the other prevents the model from inventing a new shot.

The extra line about color grading, clothing, and environment is treated as a quality control measure. Wilcock shows a side-by-side comparison with and without that instruction and says the difference can be subtle, but the added line helps when an output does not fully feel as if it belongs to the requested style. The target is not a video with a claymation coating; it is a shot that appears to have been made as claymation while retaining the original motion and framing.

This is the useful mental model for the rest of the tutorial: Gemini Omni Flash can be asked to alter the rendered content of footage, but the stronger prompts tell it where not to improvise.

Lighting is framed as a scene-level edit, not a cosmetic grade. Wilcock says a flat A-roll clip can be pushed toward the look and lighting of a professional documentary, and demonstrates a daytime street clip changed with the simple prompt “Make it night time.”

The important detail is what changes beyond the sky. In the generated result, Wilcock says the clip preserves the same elements, motion, and movement, but the environment is made consistent with nighttime. He points to background flats appearing lit and lamp posts turning on. The example is meant to show contextual inference: if the stated condition is night, then secondary objects in the scene should behave as if it is night.

The practical use case is straightforward. If footage was shot with poor lighting or at the wrong time of day, the model can be prompted to re-render the clip toward a different lighting condition while maintaining the underlying action of the shot.

Object editing is described broadly: add something, remove an unwanted element, replace a prop, change clothing, or alter background objects. Wilcock compares the precision to Nano Banana Pro’s ability to change specific elements inside an image, then says Gemini Omni Flash applies that kind of targeted editing inside video.

The street example uses the prompt “Turn the cars into red London buses.” Wilcock emphasizes that the cars are background elements, partly behind him, trees, and lampposts. In the side-by-side result, he says the cars are gone and replaced with London buses, with the occlusions respected.

He extends the same principle to more speculative edits: changing the whole apparent location so he seems to be walking through New York, London, or “the middle of nowhere”; changing his outfit into a bright pink suit; replacing a peach held by Uni with a banana, apple, or watermelon; or changing water poured into a glass into orange juice or lava.

The glass-hand example is the most technically revealing. Wilcock says he prompted the model to turn his hand into glass, and argues that Gemini Omni Flash handles transparent material well. When the glass hand passes in front of his face, his facial features remain visible and recognizable behind it.

That claim is the basis for Wilcock’s explanation of character consistency: the model is not just continuing from a single supplied starting frame; it is using the full video as context for the edit.

After modifying the underlying footage, Wilcock turns to elements added on top of it: subtitles, titles, motion graphics, and full layouts. Subtitles can be generated from the speech in the video, styled and animated by prompt, without manually designing or timing them.

The subtitle prompt shown asks for “vintage typewriter subtitles” at the bottom of the screen, based on the video’s audio, synced perfectly to the timing of the spoken words. Wilcock says adding that sync instruction helps produce more consistent outputs. He also gives the caveat that matters in production: if the model hallucinates subtitles, struggles with timing, or gets words wrong, the user may need to type the exact subtitle text into the prompt. That can reduce the number of failed generations.

Text titles follow the same promptable approach but benefit more from design vocabulary. A simple prompt asks for “a claymation title in big behind the man that says ‘Alec Wilcock.’” Wilcock also suggests prompts for lower thirds, titles behind the subject, and eased-in animations. The more useful lesson is that traditional editing terminology improves results: specify how the title enters, how it moves, what material it resembles, where it sits, and how it interacts with the scene.

The cheat sheet also includes typography terms such as heavy or bold weight, outlined type, wide letter spacing, and lower third; material references such as clear glass with refraction, inflatable glossy vinyl, paper cutout edge, sticker with a white stroke, chalk or dry erase, and neon tube glow; and polish instructions such as adding a matching sound effect or making only one element slow motion rather than the whole clip.

The most expansive workflow is a vertical selfie video turned into a TikTok or Instagram Reel. Wilcock says Gemini Omni Flash can add motion graphics, modify the layout, apply color grading, create dynamic subtitles, and build a more complete short-form edit from one detailed prompt.

The prompt shown is highly constrained. It tells the model to keep the speaker fully photorealistic and unchanged: same face, skin texture, clothing, movement, timing, gestures, voice, and audio. It instructs that the original speech remain untouched and clearly audible above added sound. The layout brief then rebuilds the frame as a modern interface-style motion design piece: the original footage becomes a rounded-corner card with a drop shadow, off-white border, and soft diffused shadow on an off-white background with a barely visible dot grid.

Around that card, the prompt asks for huge black display typography showing the key word of each sentence, white pill-shaped tags that stack as points are made, a progress bar, and an occasional cursor click that flips a pill to lime green. It limits the palette to lime green and black on off-white, requires grid alignment and consistent spacing, forbids rotated or scattered elements, specifies 250-millisecond ease-in-out motion with no overshoot, bans emojis, and says never to cover the speaker’s face. It also asks for a low-volume electronic bed, UI pop sounds, and a whoosh when the card changes position.

Wilcock’s practical warning is that vague intent is not enough. Simply asking to “turn this into an engaging Reel” will not produce the same quality of result. The more detailed the brief — style, palette, motion, layout, negative constraints, and sound — the more distinctive the output can become. He suggests using an LLM such as ChatGPT or Claude to help draft that prompt if needed, and notes that ElevenCreative can generate audio, images, and video in the same environment.

A final workflow starts with an image rather than footage: an ElevenCreative logo placed in the image reference slot. Because the source is static, Wilcock says the user can choose duration, number of generations, and aspect ratio. His example keeps a 16:9 aspect ratio for horizontal content and mentions a five-second animation as a possible duration.

The prompt structure is simple but not generic: “Animate,” tag the asset as “@image 1,” start with a blank frame, then describe the motion. His example asks for the first half of the logo to swipe in from the left and the remaining letters to fall from the top. He adds “Add sound effects,” saying Gemini Omni Flash will generate sounds to accompany the motion.

Across the workflows, the same constraint holds: the model can restyle, relight, replace, subtitle, title, lay out, animate, and add sound, but the useful prompts behave like edit briefs. They specify the intended change, the assets or features that must be preserved, the motion language, the scene integration, and the boundaries where the model should not invent.