loading
loading
Artificial Analysis Ranks Seedance 2.5 and MiniMax H3 in Video AI Benchmark Artificial Analysis just rolled out AA-Video-I2V v2.1, a 1080p image-to-video benchmark with 58,000+ blind human votes across 10 use cases and 10 capabilities. - Artificial Analysis launched AA-Video-I2V v2.1, a 1080p image-to-video benchmark with per-category rankings. - 58,000+ blind votes on 1,000 prompts for the audio board, 44,000+ votes on 500 prompts for silent. - Dreamina Seedance 2.5 ranks #1 overall and leads 11 category boards at $34.12/min. - Open-weights MiniMax H3 (768p) hits #3 at $4.80/min, sitting on the price-quality Pareto frontier. - Only 5% of prompts are public and the set rotates monthly to resist benchmark contamination. - Text-to-video winners often lose at image-to-video: Wan 3.0 drops from #1 on T2V to #5 here. Artificial Analysis has launched a taxonomy-driven benchmark for image-to-video models, giving developers an alternative to choosing from demo reels and social media clips. The rankings evaluate included models at output resolutions up to 1080p and break performance down by production use case, technical capability, and visual style. Two companion boards cover different output formats. AA-Video-I2V v2.1 evaluates image-to-video models that generate synchronized audio, while AA-Video-I2V-Silent v2.1 covers video without audio. Both use blinded, pairwise human preference votes. Each test combines a 1080p source image with a prompt written for that image, then assigns tags for its use case, capability, and style. Hidden prompts and checked votes The audio benchmark debuted with more than 58,000 votes across 1,000 prompts. The silent board collected more than 44,000 votes across 500 prompts. Model entries can run below the 1080p ceiling, as the MiniMax H3 768p listing illustrates. Only 5% of the prompts appear in the public arena, and Artificial Analysis rotates that subset monthly. Limiting exposure reduces benchmark contamination, where developers optimize models against known test prompts instead of improving broader performance. Each prompt receives three tags that determine the category boards and balance the benchmark across common production needs: - 10 use cases: Marketing & Advertising, Retail & Ecommerce, Live-Action Film, Animation & Gaming, Architecture & Real Estate, UI/UX & Motion Design, and four additional categories. - 10 capabilities: Human Anatomy, Camera Control, Text Rendering, Spatio-Temporal Consistency, Lighting & Materials, Audio Synchronization, Dialogue & Lip Sync, and three additional categories. - 8 visual styles: including Photorealistic, 3D Render/CGI, Cartoon & Anime, and Flat Design. Matchups are sampled evenly across use case and capability combinations, which gives breadth across categories more weight in the overall score. Under the published methodology, evaluators watch two videos generated from the same image and prompt, with model identities hidden, and choose the output they prefer. Voting unlocks only after both videos have played for at least eight seconds. Roughly one in five matchups is a quality-control test with an established consensus answer. Artificial Analysis discards sessions that fail those checks. The resulting scores measure human preference under this prompt set, so close rankings can reflect sampling uncertainty rather than a dependable quality gap. Seedance leads; MiniMax cuts the cost Dreamina Seedance 2.5 ranks third on the audio leaderboard with an Elo score of 1138. Its wins include Human Anatomy, Camera Control, and Retail & Ecommerce. At $34.12 per generated minute, it is also the most expensive model in the audio top 10. MiniMax H3 at 768p ranks third while costing $4.80 per generated minute, about one-seventh of Seedance 2.5’s listed price, and leads the Text Rendering category. It also sits on the price-versus-quality Pareto frontier, meaning no tested alternative is both cheaper and more preferred. The two audio models above it, Seedance 2.0 and Seedance 2.5, cost about five and seven times more, respectively. MiniMax H3 Max leads the silent board, while Video Rebirth’s Bach 1.0 Pro is statistically tied with it at $7.20 per generated minute. The voting data therefore does not reliably separate those two models at the top. A fixed first frame reshuffles the field Animating a supplied image adds a preservation constraint that text-to-video generation does not face. The model must introduce motion while retaining the people, products, composition, text, and visual style already present in the frame. Results from the two benchmarks show that strong text-to-video performance does not consistently predict image-to-video performance. - Wan 3.0 leads the AA-Video-T2V v2.1 text-to-video board but falls to fifth on image-to-video. Its largest gaps appear in Dialogue & Lip Sync, Lighting & Materials, and Spatio-Temporal Consistency. - Dreamina Seedance 2.5 ranks fifth on text-to-video and first on image-to-video. It is also statistically tied for the lead in maintaining scene consistency during motion. - Gemini Omni Flash 1.1 leads UI/UX & Motion Design on text-to-video but ranks sixth when animating a source image. It retains strong Audio Synchronization results in both formats. - Dreamina Seedance 2.5 remains the photorealistic leader across both boards. For Cartoon & Anime prompts, it rises from seventh in text-to-video to first in image-to-video, where the source frame anchors the style. Choose by workload, then test locally Teams building from product photography, storyboard frames, interface mockups, or character designs can use the category boards to narrow the field before running internal evaluations. The overall ranking rewards versatility, while capability scores expose strengths that matter to specific pipelines, such as text preservation, camera motion, lip synchronization, or temporal stability. - Product and ecommerce video: compare Retail & Ecommerce, Text Rendering, Lighting & Materials, and Spatio-Temporal Consistency. - Dialogue and character work: prioritize Human Anatomy, Dialogue & Lip Sync, and Audio Synchronization. - Interface animation: inspect UI/UX & Motion Design and Text Rendering instead of relying on the overall rank. - Self-hosted deployment: MiniMax H3’s open weights provide an alternative to a closed inference API, subject to license, hardware, and serving requirements. Because the benchmark centers on output preference and listed generation price, production evaluations should separately measure latency, throughput, repeatability across seeds, API reliability, moderation limits, licensing terms, and performance on proprietary images. Prices also require verification against current provider rates and the resolution used in deployment.
document fetched 2026-10-29 03:02