Artificial Analysis launches image-to-video benchmark: Seedance 2.5 tops, MiniMax H3 best value

ArtificialAnlys · x · 2026-10-29

Artificial Analysis released AA-Video-I2V v2.1 and a silent variant, ranking image-to-video models at up to 1080p across 10 use cases, 10 capabilities and multiple visual styles, based on 58,000+ (44,000+ for silent) human preference votes over 1,000+ prompts, with only 5% of prompts public and monthly rotation to resist contamination.

Key findings:

Dreamina Seedance 2.5 ranks #1 overall, statistically tied with Seedance 2.0 and the MiniMax H3 family; it leads 11 category boards (Human Anatomy, Camera Control, Retail & Ecommerce) but is the priciest in the top 10 at $34.12/min.
MiniMax H3 (768p), open weights at $4.80/min (1/7 of Seedance 2.5's price), ranks #3, leads Text Rendering, and sits on the quality-price Pareto frontier of both benchmarks.
MiniMax H3 Max (post-trained by fal) tops the silent leaderboard and leads UI/UX & Motion Design and Spatio-Temporal Consistency; Bach 1.0 Pro ($7.20/min) is statistically tied for #1 there.
Image-to-video is a different job from text-to-video: Seedance 2.5 is only #5 on T2V but #1 here, while Wan 3.0 leads T2V yet drops to #5, weakest at lip sync, lighting/materials and consistency.
On Audio Synchronization, Seedance 2.0 leads; Gemini Omni Flash 1.1 is #9 overall but #3 in both modalities.

Original post →

Artificial Analysis launches image-to-video benchmark: Seedance 2.5 tops, MiniMax H3 best value

More from Multimodal