Seedance 2 vs Google Omni vs Runway Aleph 2

Before we get into which AI video tool won the head-to-head, let me flag something the filmmaker CyberJungle buries near the end of his 22-minute test—because it's actually the most important finding in the whole piece, and not for the reason he thinks.

He couldn't complete several tests with Google's model because he kept hitting content policy rejections he couldn't explain. A robot-chase scene, apparently reminiscent of Star Wars characters. A green-screen performance. A crowd scene with flags in the background. Each time: violation error, no further context. His read on it was frustration. Mine is: that's a production dependency you need to understand before you spend a single dollar building a workflow on this tool.

I'll come back to that. First, the actual test.

A note on product names: CyberJungle refers to the Google tool throughout as "Google Omni." This appears to reference Google's video generation capability accessed through its AI platforms—likely involving Veo 3 or a similar model integrated into Google's ecosystem. "Google Omni" isn't a clearly established standalone product name, and Google hasn't used that branding publicly. I've preserved the name he uses for readability, but treat it as shorthand for whatever Google is currently calling its video editing AI, not a confirmed product label.

Similarly, he references "GPT Image 2" as part of Runway Aleph's keyframe workflow. This may refer to GPT-4o's image generation capabilities or another model entirely—the naming in this space moves faster than the documentation.

What he actually tested

CyberJungle spent thousands of credits across all three platforms running tests that matter to working filmmakers: cinematic VFX, camera angle changes, character swapping, adding AI characters to real-world footage, storyboard-to-video conversion, and commercial applications like recoloring a car for different markets. Not benchmarks designed to make a press release look good. Actual production tasks.

Runway Aleph 2 exits the competition early. It's not really competing in the same category. Its UX is built around keyframe editing—you pick a frame, modify it using an image generation model, then let the Aleph engine interpolate the full video. That's a legitimate and potentially powerful workflow for certain use cases, but it's architecturally different from what the other two do. The face consistency problems he documents—characters changing appearance dramatically between a wide shot and a drone shot, robot serial numbers flickering between F1 and F11—aren't just polish issues. They're a sign that coherent identity across a full video sequence isn't what Aleph 2 is optimized for. If you need precise keyframe control and you're comfortable with that constraint, there's a version of this tool that works for you. If you need a character to stay themselves across an entire scene, look elsewhere.

That leaves Google's model versus Seedance 2.0, and the findings split cleanly along two axes: intelligence and reliability.

Where Google's tool wins

The "intelligent edit prompts" section is where CyberJungle's enthusiasm is most justified. He writes a prompt telling the model to swap a new person into the frame every second, specifying a rapid-fire pace and "whimsical fashion and accessory items" without naming any specific items. The model fills in the specifics coherently. Then he pushes further: "change her head every time she makes a micro face gesture." The model tracks facial movement and triggers the swap accordingly.

That's not a party trick. That's contextual understanding of human motion applied to video editing, and it's genuinely different from prompt-following in the conventional sense.

The visual intelligence tests reinforce this. Asked to replace a single ice cream with multiple flavors, Google's model not only adds the right number of flavors but sequences them in the exact order specified in the prompt. Asked to add text to existing footage, it renders the text with proper surface integration. The model carries world knowledge into the editing task in ways Seedance 2.0 doesn't quite match.

For B-roll generation, quick commercial variants, and prompt-driven creative experiments, Google's speed advantage also matters. CyberJungle calls it the fastest tool on the market for video editing, and the agent integration—where you can attach a video and ask for five different camera angles simultaneously as separate outputs—is a genuine workflow accelerator.

Where Seedance 2.0 wins

Physics. Consistency. Cinematic control.

In the spaceship-landing-in-a-bazaar test, Seedance 2.0 produces the most physically believable result: dust spreads realistically around the landing craft, the environment responds. Google's model produces a more atmospherically integrated result but the bystanders keep shopping as a spaceship lands on them, which is the kind of detail that breaks immersion in actual footage.

Character consistency is where the gap is most production-relevant. When CyberJungle swaps characters using a reference sheet, Seedance 2.0 holds the character's appearance across the full sequence. Google's model occasionally "brushes" facial features, softening realistic detail. For a 10-second social media cutaway, that's acceptable. For a scene that needs to cut with other footage, it's not.

Seedance 2.0 also handles camera movement with what CyberJungle calls "fantastic prompt following"—push-in, dolly-out, crane shots, all executed cleanly. And at 15 seconds versus Google's 10-second limit, even a five-second difference in maximum clip length changes how a sequence feels. Rushed versus grounded, as he puts it. That's not a minor spec difference; it's the difference between a usable clip and one that needs to be patched.

The black box problem

Here's what I want to spend a moment on, because it's the part of this test that matters most to anyone making a real resource commitment.

Google's model blocked CyberJungle on multiple occasions with no actionable explanation. A robot chase scene—blocked. A green screen performance—blocked. A street crowding prompt, apparently because flags were visible in the footage—blocked. His summary: "I feel like a really great model becomes unusable because of heavy censorship. The limitations are making this model sometimes unusable and this is so difficult because you don't even know sometimes what's the reasoning behind."

I spend a lot of time writing about what happens when people build workflows on systems they can't fully predict or control. Content moderation policy as a black box is a well-documented problem in social media and cloud services, and it's arriving in AI creative tools with the same structural issues: opaque criteria, no appeals mechanism described, no way to know in advance which prompts will fail. If you're building a production pipeline—commercial videography, branded content, client work with deadlines—a tool that randomly refuses tasks with no diagnostic information is not just frustrating. It's a reliability risk. You can't build a professional dependency on a system whose failure mode is "try to figure out what you did wrong."

Seedance 2.0, by contrast, processed every task CyberJungle threw at it without triggering safety filters once. That's not a small advantage for working creators. That's the difference between a tool you can plan around and one you're negotiating with.

So which one should you actually use

If someone asked me over coffee, here's where I'd land:

Use Google's model for B-roll generation, commercial color variants, intelligent text overlays, and any task where you're doing quick, AI-driven experimentation. The agent feature alone—spinning up five camera angle variants from a single prompt—is legitimately useful for pre-production. Just don't build anything client-facing on it until Google provides clearer guidance on what its content policy actually prohibits, because right now that's unknowable.

Use Seedance 2.0 when character consistency, physics, and cinematic control are load-bearing. Anything where a character needs to hold their identity across a full scene, anything where the footage needs to cut cleanly with real-world material, anything where a client will watch it frame by frame. CyberJungle notes that Seedance 2.0 costs more per generation than Google's tool on a comparable account tier—he was able to generate around 33 videos on his Google Pro account, which he found limiting, and suggests Google's pricing advantage is significant—but the pricing comparison is his, based on his specific usage, and will vary based on your generation settings and account type. Do your own math for your own workload.

Use Runway Aleph 2 if your workflow centers on precise keyframe manipulation and you want image generation tools bundled in. Don't use it as a primary video generation engine when character coherence matters.

The fact that no single tool dominates across every task isn't a hedge—it's the actual finding. These aren't interchangeable products. They're different bets on what AI video editing is for. The question worth asking before you subscribe to any of them is: what specifically do I need this to do, and can I afford it when it says no?

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.