Head to head: Kandinsky5 vs Wan v2.6 Image to Video
Kandinsky5 vs Wan v2.6 Image to Video
By Ryan Merket · Published · Updated
Four demanding video prompts put camera movement, physical action, and mood fidelity under the microscope. The comparison turns on trade-offs between making the requested action legible and keeping a shot coherent—not on a consistent advantage for either model.
Wan v2.6 Image to Video edges the aggregate, 26.4 to Kandinsky5’s 25.6, but that gap is too small to carry a verdict: confidence that either model is genuinely better is just 50%. The task-level assessments also change direction across judge passes, which is a strong sign that these clips are effectively even rather than evidence of a reliable lead. The midnight track prompt captures the split neatly. Kandinsky5 tends to sell the rain-dark, amber-lit atmosphere and upward camera move; Wan can make the ladder drill and wider stadium easier to read, but risks washing out the midnight mood or making the camera movement feel less continuous. Neither consistently nails both the scene and the requested reveal. The same trade-off appears in the other tasks. On the balloon burst, one read favors Kandinsky5’s continuous expanding sheet while another finds Wan’s balloon-to-droplets transition clearer. For the cathedral, the models trade a coherent glide down the nave against a more direct, centered approach to the altar. In the wrist-wrap scene, Wan often makes the wrapping action more legible, while Kandinsky5 is praised for continuity and the attempt to show the tape being torn; neither reliably delivers every requested beat. **Final call: a sample tie. Wan’s 0.8-point aggregate edge is not a meaningful win, and the task results give no stable basis for naming a better model.**
Crane Up on Midnight Pacer
A one-shot 16:9 video of a lean middle-distance runner in a cobalt singlet doing fast-feet ladder drills on the rain-dark infield beside lane 5 of a nearly empty track, tiny amber timing LEDs blinking at his shoes; the camera performs a smooth crane up exactly as he accelerates through the last four rungs, rising from shin height to well above his head without tilting away, revealing the wet oval, drifting breath, and faint scoreboard glow under misty stadium lights, with a focused, electric pre-race mood.
OpenAI: GPT-6 Luna: Model A better preserves the dark, amber-lit mood and suggests a smoother upward camera move, but never reveals the runner or the wider track fully. Model B provides a more complete view of the ladder drill, wet oval, scoreboard glow, and stadium, though its framing shifts feel less continuous and the lighting is too bright for midnight. (Second judge pass, order swapped — scores are the average of both: Model B better captures the rain-dark track, cobalt kit, amber ladder lights, and wide stadi | Anthropic: Claude Fable 5.1: Model A reads as a single continuous crane rise…
Fluid & particle dynamics
Cinematic slow-motion of a water balloon bursting, the sheet of water expanding and droplets scattering through the air in convincing detail against a dark background, hard side light, 16:9.
OpenAI: GPT-6 Luna: Model A clearly progresses from an intact balloon to an expanding radial sheet and scattering droplets, with dramatic side lighting, though the burst is unusually symmetrical. Model B has crisp droplets and attractive lighting, but the sphere remains nearly unchanged before an abrupt cut to scattered water with little visible expanding sheet. (Second judge pass, order swapped — scores are the average of both: Model B shows a convincing transition from an intact water balloon to scattered droplet | Anthropic: Claude Fable 5.1: Model A shows a continuous burst but the explos…
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 presents a clear, centered approach toward the altar with consistent framing and visible dusty light, while Model A glides through a candlelit side view without clearly revealing the altar. Model A has rich candlelight, but Model B better matches the requested movement and maintains stronger visual continuity. (Second judge pass, order swapped — scores are the average of both: Model B maintains a centered, slow-looking glide down the cathedral aisle toward the altar, with consistent cand | Anthropic: Claude Fable 5.1: Model B frames the nave directly from the entra…
Handheld Tape Wrap Before Lift
A one-shot 16:9 video of a compact female weightlifter with a shaved undercut and chalk-dusted forearms sitting on a low wooden plyo box, carefully wrapping crimson athletic tape around her right wrist with her left hand, biting the tape to tear it clean, then flexing and rolling the wrist once to test it; the camera uses a handheld follow, staying close and subtly circling from her left shoulder to a front three-quarter view in one fluid move, under cool fluorescent gym lighting with floating chalk haze and a tense, quietly determined competition warm-up mood.
OpenAI: GPT-6 Luna: Model B more clearly shows the tape being wrapped and the wrist being tested, though neither entry includes the requested bite-to-tear and Model B’s framing and setting shift noticeably. Model A has stronger cool, hazy gym atmosphere and steadier visual continuity, but the wrapping action is less evident and both wrists appear already wrapped. (Second judge pass, order swapped — scores are the average of both: Model B more clearly shows the requested wrist-wrapping sequence, chalk-dusted arms, a | Anthropic: Claude Fable 5.1: Model B actually performs the prompted action—a…
Matchup powered by OpenRouter.