New Image-to-Video Leaderboard: The First Frame Is Already Drawn—Can the Model Keep It From Falling Apart?

Artificial Analysis has released AA-Video-I2V v2.1 and AA-Video-I2V-Silent v2.1, two benchmarks specifically designed to evaluate image-to-video models. The former includes audio, while the latter is silent. All models run a new set of image-plus-text prompts at up to 1080p and are ranked across 10 use cases, 10 capabilities, and multiple visual styles.

Image-to-video and text-to-video do two different jobs. Text-to-video generates an entire scene from a prompt; in image-to-video, the first frame already exists, and the model must make that image move without changing the people, products, style, or layout within it. Creators often start from a product image, a storyboard frame, or a character design, and their need is to make them move, not to redraw them.

Votes for this benchmark come from private evaluators in the United States and the United Kingdom: the audio version has more than 58,000 votes covering over 1,000 prompts, and the silent version has more than 44,000 votes covering over 500 prompts. Only 5% of prompts appear in the public arena, and they rotate monthly, in order to reduce public-set contamination.

The overall leader is Dreamina Seedance 2.5. It is statistically tied with Seedance 2.0 and the MiniMax H3 family, but it takes first place in 11 category leaderboards—more than any other model—including Human Anatomy, Camera Control, and Retail & Ecommerce. Its price is $34.12/min, the most expensive in the top ten.

Dreamina Seedance 2.0 ranks second, taking first place in Audio Synchronization and Lighting & Materials.

MiniMax H3 (768p) ranks third, is open-weight, and costs $4.80/min—roughly one-seventh of Seedance 2.5. It ranks first in Text Rendering.

MiniMax H3 Max, post-trained by fal, ranks fourth in the audio version and first in the silent version, leading Spatio-Temporal Consistency and UI/UX & Motion Design.

Video Rebirth's Bach 1.0 Pro ranks third in the silent version, statistically tied for first, at $7.20/min.

In the price-versus-quality comparison, MiniMax H3 (768p) is on the Pareto frontier of both leaderboards at the same time. With audio, only Seedance 2.0 and Seedance 2.5 rank above it, at roughly 5x and 7x its price. Without audio, Bach 1.0 Pro ($7.20/min) and MiniMax H3 Max ($9.60/min) also enter the frontier. Open weights give near-frontier image-to-video capability a low-cost option that does not depend on a closed-source API.

One notable phenomenon: strong text-to-video performers are not necessarily still strong in image-to-video. Dreamina Seedance 2.5 ranks 5th on AA-Video-T2V v2.1 but 1st here. It preserves the starting image best and is statistically tied for the lead in Spatio-Temporal Consistency. Wan 3.0 leads the text-to-video leaderboard but ranks 5th in image-to-video; its three largest gaps are Dialogue & Lip Sync, Lighting & Materials, and Spatio-Temporal Consistency.

Spatio-Temporal Consistency measures something very core: once the picture starts moving, can people, objects, and scenes remain consistent from frame to frame? Seedance 2.5 and MiniMax H3 Max are statistically tied for the lead, Seedance 2.0 follows closely, and Wan 3.0 ranks fourth.

Camera Control looks at whether a model can execute pans, dolly moves, orbits, or zooms on command, smoothly and on time. In image-to-video, the camera has to move outward from a fixed first frame without breaking that first frame. Seedance 2.5 leads, followed by MiniMax H3 Max and Seedance 2.0. Wan 3.0 is statistically tied for first in text-to-video Camera Control but ranks 5th here.

UI/UX & Motion Design covers app and web interface animations, product demos, animated icons, and motion graphics. The challenge here is keeping flat, precise designs exactly as they are. MiniMax H3 Max leads, and the base model MiniMax H3 (768p) ranks second. Gemini Omni Flash 1.1 ranks first in text-to-video UI/UX & Motion Design but 6th when animating from a starting image.

Retail & Ecommerce looks at product videos and e-commerce content, where the product must look exactly like it does in the photo. Seedance 2.5 leads, followed by Seedance 2.0 and Wan 3.0.

Audio Synchronization looks at whether sound effects, music, and ambient sound match the visuals, with both timing and type correct. Seedance 2.0 leads, followed by Seedance 2.5 and Gemini Omni Flash 1.1. Gemini Omni Flash 1.1 ranks only 9th overall but 3rd in this category; it also leads text-to-video Audio Synchronization, making sound a strength for it in both modalities. MiniMax H3 Max is the opposite case: it ranks 1st without audio and 4th with audio, and Audio Synchronization is precisely its weakest category.

Style also changes the rankings. Under realistic prompts, Seedance 2.5 ranks first in both text-to-video and image-to-video. Stylized categories are different. When the style comes from the starting image rather than the prompt, Seedance 2.5 goes from 7th in text-to-video to 1st in image-to-video on Cartoon & Anime. Wan 3.0 ranks 2nd in text-to-video Cartoon & Anime but 5th here. MiniMax H3 Max leads Flat Design.

Beyond the leaderboard, one question is worth noting. Commenter NeuralLoot raised the point that 58,000 preference votes may reward smooth motion rather than fidelity to the first frame. He asked whether raters view the starting image and the generated result side by side, so that deviating from the first frame would actually cost a model. The leaderboard does not directly answer this question, but it points to the hardest part of image-to-video evaluation: good-looking motion is one thing; whether the people, products, and layout from the first frame have been preserved is another.

Full results are here: https://artificialanalysis.ai/video/leaderboard/image-to-video

Methodology: https://artificialanalysis.ai/video/methodology