Skip to main content
Back

Artificial Analysis Launches a Video Editing Arena Pitting 6 Frontier Models

Artificial Analysis launches a blind-vote Video Editing Arena pitting Seedance 2.0, Runway Aleph 2.0, Wan 2.7, HappyHorse 1.0, Kling 3.0 Omni, and SkyReels V4 head-to-head on real editing tasks

Paper, Full Access
  • New arena launched: Artificial Analysis opens a Video Editing Arena with blind human-preference voting across 6 frontier models.
  • Six models competing: Seedance 2.0, Runway Aleph 2.0, Wan 2.7, HappyHorse 1.0, Kling 3.0 Omni, and SkyReels V4 are all in the initial lineup.
  • Editing-specific categories: Tests focus on VFX, sound/speech editing, object removal/addition, restyling, relighting, and physics simulation.
  • Elo ranking system: Models are scored via blind pairwise comparisons; first leaderboard results expected within 24 hours of launch.
  • Community submissions open: Anyone can submit a 5-10s clip they own with an editing instruction to be tested across all models.
  • No single model dominates editing yet: Early analysis shows clear differences in model strengths by editing category, making the benchmark practically useful for production decisions.

Video generation benchmarks have been around for a while, but benchmarking video editing , taking an existing clip and modifying it with a text instruction , is a different problem entirely. Artificial Analysis just launched a dedicated Video Editing Arena to fill that gap, putting six of the most capable frontier models through blind, human-preference comparisons on real editing tasks.

What the Arena actually tests

The arena is structured as a head-to-head bracket: you watch two edited versions of the same source clip, without knowing which model produced which, and vote for the better result. Rankings are derived from an Elo rating system built on user votes in blind comparisons, where users compare two outputs from the same input without knowing which model created each one. Higher Elo scores mean a model wins more often.

The editing categories are deliberately chosen to reflect where models actually differ in practice:

  • Visual effects (VFX) editing
  • Sound and speech editing
  • Object addition, removal, and replacement
  • Restyling and relighting
  • Background changes
  • Physics simulation

Crucially, the arena tests both with-audio and no-audio conditions , and when audio is included, you're judging the sound too, so headphones in a quiet space are recommended. This matters because audio-visual synchronization is one of the sharpest differentiators between today's frontier models.

The six models in the ring

The launch lineup covers the main contenders in the video editing space right now:

  • Dreamina Seedance 2.0 , ByteDance's multimodal model. It accepts images, videos, audio, and text as inputs, giving creators control over every aspect of generation. It preserves product details, logos, and text across frames, which matters for e-commerce.
  • Runway Aleph 2.0 , Runway's flagship video editing model, built for precise localized edits, single-frame visual guidance, and consistent edits across cuts and scene changes. It supports clips up to 30 seconds at 1080p.
  • Wan 2.7 , Alibaba's next-generation multimodal model supporting text-to-video, image-to-video, and video editing, with features including style transfer, instruction-based editing, and audio-conditioned generation with auto-dubbing.
  • HappyHorse 1.0 , A newer entrant that has already made a strong showing on Artificial Analysis's existing text-to-video leaderboard. It may lead the leaderboard, but without a public API or usable interface, it remains out of reach for most production workflows.
  • Kling 3.0 Omni , The Omni variant unifies video, audio, image, and editing in one architecture, with native lip-sync in five languages and multi-shot storyboards up to 6 shots per clip.
  • SkyReels V4 , A unified multi-modal video foundation model that jointly generates video and audio while supporting generation, inpainting, and editing within a single architecture, producing outputs at up to 1080p, 32 FPS, and 15 seconds.

Why editing is harder than generation

Generating a video from scratch and editing an existing one are fundamentally different tasks. In editing, the model has to follow a specific instruction while keeping the rest of the clip stable , preserving motion, lighting, audio sync, and identity across frames that it didn't generate. Recent efforts like Kling 3.0, Seedance 2.0, and Vidu-Q3 have taken steps toward joint video-audio generation, but the field still lacks a unified foundation model capable of seamlessly handling generation, inpainting, and editing conditioned on arbitrary combinations of text, images, videos, masks, and audio references.

That fragmentation is exactly what makes a dedicated editing arena useful. The video editing space is evolving, but still fragmented across providers , no single model dominates yet. Different models have clear strengths by editing type: Runway Aleph 2.0 is purpose-built for localized, temporally consistent edits; Seedance 2.0 leads on audio reference control; SkyReels V4 handles inpainting natively within its architecture.

How the existing leaderboards set the stage

Before the editing arena launched, Artificial Analysis's text-to-video and image-to-video leaderboards already showed a competitive field. Dreamina Seedance 2.0 currently leads among text-to-video models with audio output with an Elo score of 1218. HappyHorse 1.0 leads the text-to-video arena without audio at an Elo score of 1290. The editing arena is a separate category , a model's generation ranking doesn't automatically predict its editing ranking.

That gap between generation and editing quality is the key insight here. The video editing arena specifically evaluates instruction-based modifications to existing clips , style transfer, object removal, and scene changes , tasks that require a different kind of model capability than generating from a blank slate. Early analysis from Artificial Analysis already reveals clear differences in model strengths by editing category.

What's open and what's not

The arena is free to use and open for voting now at artificialanalysis.ai/video/arena. The first leaderboard results are expected within 24 hours of launch. Artificial Analysis is also accepting community prompt submissions , you can submit a source clip with audio and an editing instruction to have it tested across all six models. Requirements:

  1. Short clip (~5-10 seconds) that you own the IP for
  2. Safe for work (submissions are moderated)
  3. A real editing use case: VFX, SFX, object removal/addition/replacement, restyling, relighting, background change

Submit by replying to the original thread with your clip and instruction.

Why this benchmark matters now

The practical stakes are real. For post-production workflows involving style changes or scene edits, the video editing arena rankings are the most relevant signal. VFX artists, motion designers, and content teams are already using these models in production, and the lack of a rigorous, human-preference-based editing benchmark has made model selection largely guesswork.

The Artificial Analysis Video Arena measures one thing: which video real users preferred in a blind comparison. It tells you nothing about API stability, pricing predictability, inference latency at scale, or whether you can integrate the model today , so treat the Elo scores as a quality signal, not a deployment decision. That said, having a clean, apples-to-apples quality ranking for editing tasks is a meaningful step forward for anyone building on top of these models.