Artificial Analysis发布视频编辑排行榜

HappyHorse-1.0 from @HappyHorseATH takes #1 in AI video editing, followed by Seedance 2.0 from @dreamina_ai and Wan 2.7 from @Alibaba_Wan, on the new Artificial Analysis Video Editing Leaderboard 🏆 In previous media generation modalities, we focussed on overall model quality. In Video Editing, we have started to identify the drivers of overall model quality. We score models across five editing capabilities, each reflecting how these models are actually used in the real world: ➤ Visual Effects Editing: VFX, relighting, background change, and restyling, valuable for creative production such as film and advertising ➤ Physics & World Simulation: how a scene moves through space and time under real or even fictional physics, valuable for creative direction and storytelling ➤ Sound & Speech Editing: sound design and dialogue, valuable for shaping mood and story or localizing narrated video content ➤ Object Editing: add, replace, or remove objects, valuable across ecommerce and entertainment for variants and spot fixes ➤ Complex Edits: multi-edit instructions, valuable when both controllability and efficiency matter After ~80K blind human votes, strengths are split: different models lead different capabilities. Kling 3.0 ranks #5 overall but #1 in Visual Effects Editing. In Sound & Speech Editing, Wan 2.7 rises to #2 (#3 overall), while Seedance 2.0 (#2 overall) sits at #3. Seedance 2.0 and HappyHorse-1.0 are tied at the top of Physics & World Simulation. HappyHorse-1.0 wins overall by being the only model never weak in any editing capability: top-3 in all five capabilities, and #1 or tied #1 in four of them. Full capability breakdowns and example videos below 🧵 https://video.twimg.com/tweet_video/HLrYXBbasAAkS-h.mp4 Artificial Analysis (@ArtificialAnlys): Zooming out, two patterns stand out: 1. Sound & Speech Editing is the single biggest separator, with the widest score spread of any capability. HappyHorse-1.0 and Wan 2.7 are far ahead of the field. 2. Models strongest at changing how a scene moves and sounds (Sound & Speech Editing, Physics & World Simulation, Complex Edits) tend to be the weakest at changing how it looks (Visual Effects Editing), and vice versa. Wan 2.7 is the sharpest case: #2 in Sound & Speech Editing and Complex Edits, but #6 in Visual Effects Editing. Its mirror image is Kling 3.0: #1 in Visual Effects Editing, but #6 in Sound & Speech Editing. HappyHorse-1.0 is the only model that handles both. Artificial Analysis (@ArtificialAnlys): [Capability 5/5] Complex Edits Real creative work is iterative and messy. Great videos come from back-and-forth editing, but each new generation adds friction and breaks creative flow. This series of tests packs three or more edits into one prompt. Valuable when both controllability and efficiency matter. HappyHorse-1.0 leads, with Wan 2.7, Seedance 2.0, and Aleph 2.0 close behind. HappyHorse-1.0 is the only model not weak on any edit capability, and handles complex edits the best too.