The Nano Banana Pro of Video Is Still Up for Grabs: Omni Flash vs Seedance 2.5

TL;DR

For a year the phrase that mattered in image editing was Nano Banana. Google's Nano Banana Pro (Gemini 3 Pro Image) made editing a picture feel like a conversation: point at any part of the frame, say what to change, keep everything else. Now every video lab wants that same point-and-talk magic for footage, and in the span of two weeks two contenders staked their claims. Google shipped Gemini Omni Flash, which edits a clip by talking to it turn after turn. ByteDance is rolling out Seedance 2.5, which lets you change one region of a clip and leave the rest frame-perfect. They have split Nano Banana Pro's two superpowers between them, and as of today no video model owns both. The "Nano Banana Pro of video" is still up for grabs.

What the crown actually is

Nano Banana Pro won images by fusing two things most editors treat as separate. It is conversational: you refine over many turns and it remembers the last state. And it is localized: you can select any region, an object, a face, the sky, and change just that, adjusting camera angle, focus, color grade, or lighting without disturbing the rest. Point and talk.

Video is a harder problem, and not by a little. Editing a photo means fixing one frame. Editing a clip means making that fix hold across every one of hundreds of frames while the subject walks, the light shifts, and the camera pans. Change a jacket from black to red and it has to stay red, in shadow and in sunlight, for the full length of the shot, without the person's hand smearing into the collar. A photo edit is retouching a single Polaroid. A video edit is retouching a flipbook and needing the same correction redrawn, consistently, on every page. That temporal-consistency tax is why "just do Nano Banana Pro for video" took an extra year to show up.

Gemini Omni Flash: the conversation half

Google announced Omni Flash on June 30 as the fast, cheap first member of the Gemini Omni family (we covered the launch-day economics here). Its editing pitch is conversational: generate a clip, then issue instructions that build on each other. "Change the background to a rain-soaked Tokyo street." Then "add warm amber lighting." Then "drop a glowing logo on the product." Each instruction lands on the current state of the video, and per Google the characters stay consistent, the physics hold up, and the scene remembers what came before. It is stateful under the hood: the API threads edits by a prior-interaction id, so you are not re-uploading the clip every turn.

Two things make it attractive beyond the demo. It is grounded in Gemini's reasoning, so it tends to understand what you mean rather than only what you typed, and it is cheap and everywhere: ten cents per second of output, rolling out through the Gemini app, Google Flow, and at no cost inside YouTube Shorts and the YouTube Create app. Every clip carries an invisible SynthID watermark.

Here is the catch, and it matters for the crown. Omni Flash does not do true masking, inpainting, or targeted object removal. Google's own docs say simple prompts work best, and that the model edits by preserving the elements you did not mention rather than by letting you draw a box around the one you did. It is a director you brief, not a scalpel you aim. Ask it to make the scene night and it repaints the whole thing beautifully. Ask it to swap exactly this can of soda for the client's and hold the actor's hand motion to the pixel, and you are hoping, not selecting.

Seedance 2.0 to 2.5: the precision half

ByteDance came at the crown from the opposite corner. Seedance 2.0, launched February 10 and now powering CapCut and Dreamina, is not a quiet model: it sits at the top of the Artificial Analysis video arena, an Elo of 1,219 for text-to-video, ahead of Kling 3.0 at 1,105 and Google's Veo 3.1 at 1,094. It already does targeted edits to specified clips, characters, and actions, plus scene extension, and takes an @-reference stack of up to nine images, three video clips, and three audio files.

Seedance 2.5, whose API reaches BytePlus around July 16, pushes the precision angle harder (our preview coverage is here). It generates a native 30-second clip in a single pass instead of stitching, accepts as many as 50 references, and reportedly outputs 4K. But the feature aimed straight at Nano Banana Pro's heart is localized editing: select a region and swap a product, a background, or a subject, or fix a single detail such as a character's hair color, while the original motion, camera, lighting, and composition stay locked. That is the video echo of Nano Banana Pro's "select any part of the frame."

The precision comes with its own asterisks. Seedance is pricier than Omni Flash (2.0 runs roughly 30 cents per second at 720p and 68 cents at 1080p through third parties, and 2.5 pricing is unannounced), the region-level control was shown in a demo before it was fully documented, and 2.5's launch arrived under what one outlet politely called a copyright cloud. A model that can regenerate any patch of footage is powerful. It is also the kind of power that keeps lawyers employed.

Who wins which edit

Strip away the branding and there are two different jobs hiding inside "edit the video."

• Art-direct the whole vibe, iteratively. Make it night, add rain, warm the light, restyle the grade. This is Omni Flash's home turf: talk, watch, talk again, at ten cents a second, inside apps a billion people already open.
• Change exactly one thing and lock everything else. Replace the drink, fix the hair, swap the logo, keep the take. This is Seedance 2.5's pitch: point at the region, keep the frame.

Nano Banana Pro did both in one model for stills. For video you currently pick a corner. Omni Flash has the conversation, the price, and the distribution. Seedance has the precision, the length, and the quality crown.

So who takes the crown

The honest answer is nobody yet, and that is the interesting part. The real "Nano Banana Pro of video" will be whichever model first fuses all four traits at once: conversational and localized and long and cheap. The two obvious paths there are Google bolting proper region masking onto a future Omni Pro, or ByteDance wrapping a genuine multi-turn chat loop around Seedance's region editor. Whoever closes their missing half first owns the category, the same way Nano Banana Pro owned images by refusing to make you choose.

For now the call is simple. If you live in YouTube and Google and want to talk your way to a look, use Omni Flash. If you are a CapCut editor who needs one element changed in a hero shot and the other 29 seconds untouched, wait for Seedance 2.5. And keep the fruit jokes loaded, because Google's naming committee clearly has a standing order at the same banana stand.

Key Takeaways

• The prize is the "Nano Banana Pro of video": a model that edits footage the way Nano Banana Pro edits images, by combining multi-turn conversation with point-at-a-region localized control.
• Gemini Omni Flash owns the conversational half: build edits turn by turn, characters and audio preserved, at $0.10 per second across Gemini, Flow, and YouTube. It does not do true masking or targeted object removal, so it is a director you brief, not a scalpel you aim.
• Seedance 2.5 owns the precision half: change one region of a clip while motion, camera, and lighting stay locked, with native 30-second shots, up to 50 references, and reported 4K. Pricing is unannounced and the region control was demoed before it was documented.
• Seedance 2.0 still holds the raw-quality crown, topping the Artificial Analysis video arena ahead of Kling 3.0 and Veo 3.1.
• No video model does both halves yet. Whoever fuses conversational, localized, long, and cheap first takes the category.

Sources: Google: Introducing Gemini Omni, Gemini API docs (Omni video editing), Google Cloud (Omni Flash availability and pricing), Google: Nano Banana Pro, ByteDance Seed (models), Seedance 2.0, DigitalApplied (Seedance 2.5 specs and benchmarks), TechTimes (Seedance 2.5 launch)