Head to head: MAGI-1 (Distilled) vs Seedance 2.5 US Reference to Video
MAGI-1 (Distilled) vs Seedance 2.5 US Reference to Video
By Ryan Merket · Published
This matchup tests whether a distilled model can keep pace across fast character action, precise subject manipulation, atmospheric scene construction, and controlled camera movement. The results expose a substantial gap in both prompt adherence and temporal execution.
Seedance 2.5 US Reference to Video wins decisively: **31.1 to 17.1**, with a **limited confidence** statistical verdict. It took **11 of 12 task wins**, with no ties—a broad advantage rather than one inflated by a single standout clip. Its strongest separation came in controlled motion. Seedance produced convincing milk-pour progression into recognizable rosettas, executed substantial orbits around steaming ramen, and delivered coherent lateral reveals through the storm-lit sunroom. MAGI-1 was repeatedly too static: its latte sequences barely developed, its ramen shots showed little meaningful orbit, and its ingredients sometimes morphed between frames. Seedance also followed dense scene directions more reliably, preserving details such as the brindle greyhound, kettle, tasseled lamp, toy mouse, laundry basket, and reflective hallway. It was not flawless—the latte framing could be more overhead, and one ramen move rose in height instead of holding the requested plane—but it still produced the more coherent and complete result. MAGI-1’s lone win came in a hallway sprint where it actually showed the child running while Seedance had the subject sliding prone, a consequential action error. **Final call: Seedance 2.5 US Reference to Video is the clear winner. Its 11–1 task record, 14-point aggregate lead, and limited confidence verdict reflect materially better action, camera control, continuity, and prompt adherence.**
Hallway Sock Sprint
A short continuous shot of a nine-year-old in mismatched rocket-print socks exploding down a polished apartment hallway while chasing a runaway battery toy mouse, pumping hard and nearly skidding around a corner as a laundry basket wobbles and a paper pinwheel on a bedroom door spins from the slipstream; the camera races backward low to the floor in front of the child, maintaining a fast close tracking view with strong momentum, brief edge motion blur, and a subtle handheld bounce, lit by bright winter morning sunlight slashing through doorways and flashing across the floorboards, mood playful, chaotic, and high-energy, 16:9
OpenAI: GPT-5.6 Sol Pro: Model B better captures the low frontal tracking perspective, runaway toy mouse, mismatched socks, polished hallway, laundry basket, and energetic chase. Model A has appealing sunlight and coherent action, but the camera follows from behind, the toy is less clearly a mouse, and key details such as the corner skid and responsive props are weak. (Second judge pass, order swapped — scores are the average of both: Model B better captures the low, backward-facing hallway chase, clearly showing the ru | Anthropic: Claude Fable 5.1: Model B nails the core staging—a low camer…
Subject action
A barista's hands pouring latte art: the milk stream forms a clean rosetta in the crema with natural, fluid wrist motion, no cuts, overhead close-up, soft café light, 16:9.
OpenAI: GPT-5.6 Sol Pro: Model B shows a convincing continuous milk pour with fluid progression into a clean rosetta-like pattern, strong café lighting, and high visual fidelity, though the view is angled rather than overhead. Model A is nearly static across frames and depicts an implausible hand-held milk strand over an oversized vessel, producing a spiral rather than a rosetta. (Second judge pass, order swapped — scores are the average of both: Model B clearly shows a continuous milk pour developing into a convincing | Anthropic: Claude Fable 5.1: Model A shows hands using a small tool to d…
Sunroom Storm Drift
A short continuous shot inside a narrow third-floor sunroom where a brindle greyhound dozes on a faded ochre armchair while the room stays alive around it: basil leaves and a hanging spider plant sway in cross-breezes from a cracked window, sheer curtains breathe in and out, rain beads race down the glass, low clouds drift past neighboring rooftops, and a kettle on a sideboard sends up a thin twisting ribbon of steam; the camera makes a slow intimate lateral slide from the plant shelf toward the dog at seat height, keeping the background motion visible and natural, lit by cool late-afternoon storm light with warm lamp spill from a tasseled floor lamp, mood quiet, tender, and slightly charged, 16:9
OpenAI: GPT-5.6 Sol Pro: Model B better captures the intimate plant-shelf-to-dog slide, a dozing brindle greyhound, tasseled warm lamp, kettle, and charged storm-light atmosphere with strong visual continuity. Model A includes many room details but the dog appears dark rather than brindle, sits awkwardly facing the window, and the composition feels less tender and cinematic. (Second judge pass, order swapped — scores are the average of both: Model B closely matches the brindle greyhound dozing on an ochre armchair, inti | Anthropic: Claude Fable 5.1: Model B nails nearly every prompt element…
Camera motion control
A single smooth orbit (roughly 180 degrees) around a steaming bowl of ramen on a wooden table, the camera circling at a constant speed and height in one continuous move, warm restaurant light, 16:9.
OpenAI: GPT-5.6 Sol Pro: Model B clearly depicts a substantial, consistent orbit around a steaming ramen bowl while maintaining stable composition, warm restaurant lighting, and strong visual fidelity. Model A is temporally stable but shows very little viewpoint change, weak steam, and a less convincing ramen presentation, so it does not fulfill the requested roughly 180-degree orbit. (Second judge pass, order swapped — scores are the average of both: Model B clearly executes a substantial, smooth-looking orbit around a | Anthropic: Claude Fable 5.1: Model B executes a clear ~180° orbit: the…
Matchup powered by OpenRouter.