Head to head: AnimateDiff vs Happy Horse

Happy Horse is the winner here, and on the topline numbers it’s a rout: 34.1 to 18.3, with a 97% confidence verdict, plus a 4–0 task sweep. If you’re choosing based on the aggregate result, this is not a photo finish. AnimateDiff simply didn’t post a single task win.

Where Happy Horse earned it was in the qualities that make video models usable rather than merely interesting. On the hammer-throw clip, it delivered the stronger slow-motion impression, cleaner framing, and better overall temporal coherence, even if neither model fully nailed the prompt. On the canal sprint, it was judged to capture the blue-hour mood, low waterline tracking feel, and environmental details far better. On crowd motion, it was the only model credited with a believable Tokyo scramble-crossing rather than a muddled mass of bodies. And on temporal consistency, it kept the face, raincoat, umbrella, and tracking composition steadier.

That said, the per-task notes reveal an uncomfortable truth for anyone treating the 4–0 as absolute gospel: every single task reversed in the order-swapped judge pass. In other words, this was a decisive aggregate win for Happy Horse, but not a clean, uncontested one at the prompt-by-prompt level. The judging signal says Happy Horse is stronger overall; the audit trail says these comparisons were surprisingly sensitive to evaluation order and emphasis.

So the practical read is straightforward. If you want the model that, in this test set, more often produced polished, coherent, cinematic video, pick Happy Horse. If you’re looking for a reason to call AnimateDiff competitive here, the record doesn’t give you one: its losses were broad enough that it finished nearly sixteen points back.

Final call: Happy Horse wins decisively overall. The 97% confidence and 34.1–18.3 scoreline are too strong to argue with, even if the task-level writeups suggest the sweep was messier than the final table makes it look.

How they were tested

We ran 4 fresh video tasks, generated on the fly for this matchup so neither model could prepare in advance, and had gpt-5.4 score each one. AnimateDiff scored 18.3 to Happy Horse's 34.0.

1. Hammer throw identity lock

A single continuous 16:9 shot in smooth cinematic slow motion begins chest-high and arcs clockwise around a hammer throw athlete during one full controlled training spin on an empty municipal field at sunrise; the subject is a tall woman with a copper undercut braided into a short tail, a small crescent-shaped scar above her left eyebrow, deep green singlet with a white stitched number 58, black compression shorts, mismatched wrist tape—left wrist teal, right wrist beige—and bright chalk on her fingers, and her face, body proportions, clothing details, and identity must remain perfectly unchanged from first frame to last as the camera orbit tightens slightly, with long golden side light, faint breath in the cold air, and a tense, disciplined mood.

Winner: Happy Horse — Model B better matches the requested cinematic clockwise orbit at sunrise with stronger slow-motion feel and more coherent framing, while Model A looks more like a static discus-style pose sequence and misses several identity details. Both miss some prompt specifics, but B is more visually polished and temporally consistent overall. (Order-swapped judge pass: Model A better matches the requested sunrise municipal field, cinematic orbit, and preserves the athlete’s face and styling more consistently across frames, though it misses some specified identity details. Model B departs strongly from the prompt with a different outfit/body presentation, visible equipment/track setting, and weaker identity lock despite acceptable motion continuity.)

2. Canal sprint at blue hour

A single continuous 16:9 shot at blue hour follows a lone sprint kayaker in a matte citron racing kayak powering down a narrow canal beside a neglected brick velodrome, while the camera glides low and parallel from the waterline with a slight forward creep; behind the athlete, reeds sway in crosswind gusts, silver ripples spread and rebound off mossy stone walls, loose pennant tape flutters from rusted railings, thin chimney steam rises from canal-side workshops, and layered clouds drift steadily across the last cobalt light, all moving naturally and continuously in a cool, focused, quietly electric mood.

Winner: Happy Horse — Model B matches the prompt far better with a low waterline parallel tracking shot, blue-hour mood, neglected brick velodrome/workshop setting, steam, pennant tape, reeds, and strong cinematic continuity. Model A shows a canal kayaker but misses the specified camera angle and environment, and the kayak/paddle details and overall scene feel less faithful. (Order-swapped judge pass: Model A closely matches the prompt with a low parallel waterline view, blue-hour mood, neglected brick architecture, pennant tape, steam, reeds, and convincing kayak motion. Model B misses the specified camera setup and setting, showing a high aerial canal shot in warmer light with weaker prompt alignment despite acceptable continuity.)

3. Crowd motion

A busy Tokyo scramble crossing seen from above, dozens of pedestrians crossing in different directions, each moving independently without merging or warping into one another, overcast daylight, 16:9.

Winner: Happy Horse — Model B clearly matches the Tokyo scramble-crossing prompt with an overhead urban intersection, independent pedestrian motion, and stable realistic rendering. Model A instead shows an implausibly dense crowd filling a roadway rather than a scramble crossing, with poor adherence and muddled crowd structure. (Order-swapped judge pass: Model A clearly matches the prompt with an overhead Tokyo-style scramble crossing and independently moving pedestrians that remain distinct and coherent across frames. Model B shows an aerial view of an unnaturally dense crowd filling roads rather than a scramble crossing, with poor prompt match and muddled crowd motion.)

4. 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.

Winner: Happy Horse — Model B is more temporally consistent: the man's face, yellow raincoat, and umbrella remain stable across frames with a steadier tracking composition and cleaner realism. Model A matches the yellow umbrella idea but shows more facial ambiguity/morphing and less precise adherence to the specified raincoat-and-umbrella consistency. (Order-swapped judge pass: Model A matches the prompt much better with a yellow raincoat, steady centered tracking, and strong consistency in the man's face, coat, and umbrella across frames. Model B is visually pleasing but deviates from the prompt with a different framing/aspect feel, a yellow umbrella instead of a consistent dark one, and weaker facial clarity/consistency.)

See every prompt and the full side-by-side outputs in the interactive Head-to-Head.