Head to head: Kling Video v3 Image to Video [Pro] vs Seedance 2.5 US Reference to Video
Kling Video v3 Image to Video [Pro] vs Seedance 2.5 US Reference to Video
By Ryan Merket · Published · Updated
![RuntimeWire Head-to-Head: Head to head: Kling Video v3 Image to Video [Pro] vs Seedance 2.5 US Reference to Video](https://runtimewire.com/og/head-to-head-kling-video-v3-image-to-video-pro-vs-seedance-2-5-us-reference-to-v.16x9.jpg?v=og2-r3w3fe)
A four-scene comparison puts two image-guided video models through rainy streets, a greenhouse squall, a cathedral glide, and a creek eddy. The clips are judged on how closely they follow each brief and hold together over time.
Seedance 2.5 US Reference to Video scored **30.2** to Kling Video v3 Image to Video [Pro]’s **27.0**, and took 9 task wins to Kling’s 2, with 1 tie. But the saved editorial result is a tie: **Too close to call (not statistical)**—not a clear overall win for either model on this sample. Seedance’s edge came through most consistently in the greenhouse, where it better conveyed the warm-to-cool light shift and push toward the hanging watering can, and in the cathedral, where its forward glide and dusty candlelit mood more clearly matched the brief. It also won two of the three creek comparisons, usually better capturing the narrow, cold stream and circulating foam. Kling had its own clear moments: it kept the yellow-coated man and umbrella legible in one rainy-street comparison, and more clearly showed the ribbon above the creek’s eddy in one run. The rainy-street judgments split across the three repeats, while the creek verdict also changed—evidence that the outcome depends on the scene, not just the aggregate score. Across 3 successful repeats, the mean paired score difference (Kling minus Seedance) was **-3.116666666666667**; the Student-t 95% confidence interval for that mean difference was **-12.972413251312854 to 6.73907991797952**. This interval describes the paired mean difference, not a probability of winning. **Final call: tie; too close to call on the saved task sample.**
Temporal consistency
A man in a yellow raincoat walking toward camera down a rainy street; his face, coat, and umbrella must stay perfectly consistent with no morphing or flicker from the first frame to the last, steady tracking shot, 16:9.
OpenAI: GPT-6 Luna: Model A more directly matches the prompt, with the man’s face, yellow coat, and umbrella clearly visible and consistent as he approaches in a steady-looking shot. Model B is more atmospheric, but its face is obscured and the long rain suit differs from the specified coat. (Second judge pass, order swapped — scores are the average of both: Model B better captures the rainy nighttime street and the man’s steady approach, with a polished cinematic look; the face is mostly obscured but the coat and | Anthropic: Claude Fable 5.1: Model A keeps the man's face, yellow coat, and b…
Greenhouse squall dim
A one-shot 16:9 video clip inside a weather-beaten coastal greenhouse during a spring squall: the camera makes a slow dolly-in past wet tomato vines toward a rusted metal watering can hanging by twine, gently swaying as wind rattles the panes; over the course of the shot, sunlight streaming through the glass is gradually blotted by a fast-moving cloud so bright warm highlights slide off the leaves into cool gray dimness, then a weak strip of late sun briefly returns across the back wall, making condensation sparkle before fading again, all smooth and believable, with a calm-but-uneasy mood.
OpenAI: GPT-6 Luna: Model B more convincingly shows a freely hanging, rusted watering can among tomato vines, with a smooth push-in and a warm-to-cool lighting change that suggests the passing squall and returning strip of sun. Model A better establishes the coastal setting, but its light progression is less complete and the can appears closer to resting on the ledge. (Second judge pass, order swapped — scores are the average of both: Model B more clearly shows the requested push past tomato vines toward a twine-hu | Anthropic: Claude Fable 5.1: Model B delivers the core lighting arc of the p…
Single continuous shot
One unbroken take gliding slowly through a candlelit cathedral from the entrance toward the altar, no cuts, jumps, or transitions, dust and warm light in the air, 16:9.
OpenAI: GPT-6 Luna: Model B more clearly glides from the entrance toward the altar and better captures the requested warm, dusty candlelit atmosphere. Model A is sharper and visually consistent, but its framing changes less and feels less like a journey through the cathedral. (Second judge pass, order swapped — scores are the average of both: Model B more clearly glides from the entrance into a warm, hazy, candlelit cathedral toward the altar, with a strong sense of forward movement. Model A is visually polished an | Anthropic: Claude Fable 5.1: Model B clearly begins at the entrance doors an…
Creek foam spiral
A one-shot 16:9 video clip at blue-hour after a mountain hailstorm: the camera glides low and slowly sideways along a narrow basalt creek while a rain-soaked orange survey ribbon, snagged on a twig, whips in the current and sheds tiny beads of water; around it, tea-colored runoff curls into a tight clockwise eddy carrying foam flecks, pine needles, and three silver birch leaves that collide, separate, and get pulled under naturally, with occasional droplets splashing the lens edge, cold overcast light reflecting off the ripples, and a tense, hushed mood.
OpenAI: GPT-6 Luna: Model B shows the prompt’s defining tight eddy in tea-colored runoff, with debris and foam visibly circling and getting drawn under; Model A better conveys the cool blue-hour creek and ribbon but lacks a clear spiral. Model B is more dynamic, though its darker, tighter framing and sparse birch leaves miss some requested details. (Second judge pass, order swapped — scores are the average of both: Model B more closely captures the prompt with a pronounced tea-colored clockwise eddy, snagged orange | Anthropic: Claude Fable 5.1: Model B delivers the core of the prompt—a tight…
Matchup powered by OpenRouter.