LMArena image-to-video leaderboard: who leads now

The LMArena image-to-video leaderboard in September 2026: MiniMax H3 on top, Seedance 2.0 stuck at 1475, and where Grok, Veo 3.1 and Kling landed.

My reading of the image-to-video board on arena.ai, where the LMArena rankings live now. You'll find the original at arena.ai/leaderboard/image-to-video, and I cross-check it with the Artificial Analysis image-to-video board, which uses its own scale. Votes keep arriving after I hit publish, so trust the live pages over this snapshot.

Seedance 2.0 scored 1475 in June. On the LMArena image-to-video leaderboard it still scores 1475, and it has dropped from a shared first place to fifth anyway. Nothing about it got worse. Four models simply arrived above it.

That's the September 21, 2026 update in a sentence. It now covers 48 models and 2,081,953 votes, up from 41 models and a bit over 1.2 million when I went through it in June. What matters most to me is watching what happens to a product label once a still starts moving, which is why this board, the one where voters upload a picture and ask for motion, gets more of my attention than the text-to-video one that gets the headlines. If the voting mechanics are new to you, I walked through them in my notes on the image arena. Same system, clips instead of pictures.

The top ten on September 21

minimax-h3 (MiniMax): 1495 with 57,112 votes;
gemini-omni-1.1-flash (Google): 1488 with 3,734 votes;
wan3.0 (Alibaba): 1480 with 3,442 votes;
dreamina-seedance-2.5-720p (ByteDance): 1477 with 15,729 votes;
dreamina-seedance-2.0-720p (ByteDance): 1475 with 121,427 votes;
gemini-omni-flash (Google): 1465 with 87,155 votes;
hidream-o1-video-1.0 (HiDream): 1456, still flagged preliminary;
grok-imagine-video-1.5-720p (xAI, listed as SpaceXAI): 1456 with 137,459 votes;
flux-3-video (Black Forest Labs): 1449;
happyhorse-1.0 (Alibaba-ATH): 1442.

Look at the vote counts in second and third place. Under 4,000 each. Its own rank estimate for gemini-omni-1.1-flash stretches from 2nd to 15th, and wan3.0 could land anywhere from 3rd to 16th. I'd file both under "check back in a month."

There's a precedent for that caution, and it's sitting in sixth. In June the original gemini-omni-flash was tied with Seedance at 1475 on just 3,014 votes. It has since collected 87,155, and it settled ten points lower. I'd rather wait until wan3.0 and gemini-omni-1.1-flash each clear 50,000 or so votes before ranking them above Seedance, given how far gemini-omni-flash's own score fell once its sample caught up.

Where did MiniMax H3 come from?

MiniMax announced H3 on July 31, 2026 as one model that reads text and images in the same context as video or audio and returns clips of up to 15 seconds in 2K with sound. On the arena it leads by 7 points with a ±5 interval over 57,112 votes, and the rank estimate keeps it somewhere between 1st and 13th. That's a real lead, not a lucky week. The Artificial Analysis board agrees, with H3 Max first and the regular H3 second on its image-to-video chart with audio.

H3's weights are public, which sounds like great news until you read the H3 community license on Hugging Face. Four places are listed as excluded territories. The United States is one of them, alongside the European Union, the UK and South Korea. So if you're in Ohio and planning to run it on your own GPU, the license doesn't cover you. Nobody's tested in public yet how H3 handles packaging text or small print, so for now the license is the bigger practical caveat.

Most people met MiniMax through Hailuo. Hailuo 2.3 is 29th at 1262, with 405,786 votes behind it. Low for the table, sure. It was never built to win on overall motion. What I use it for is facial expression in close-ups, something a general vote barely rewards.

Grok isn't the surprise anymore

Back in June the eyebrow-raiser was xAI, with two Grok Imagine versions inside the top five. Its 1.5 preview was third at 1467. xAI took 1.5 out of preview on June 16 and called it its best image-to-video model yet. Today it's eighth with 1456.

Meanwhile grok-imagine-video-720p, the older one, sits 12th at 1415. Its vote count is the wild part, 554,050, far more than any other row. I suspect that says more about how often a fast, cheap model gets served in matchups than about quality. Still, the June story was "the budget option crashes the party," and the September story is quieter. Grok now lives in the top dozen like any other regular. When I want a rough take on a motion idea in seconds, Grok Imagine is where it goes, and my longer Grok Imagine write-up covers where the speed costs you.

Seedance didn't lose a point

121,427 votes and a ±7 interval make Seedance 2.0's score the most settled number in the top five. Its successor, Seedance 2.5, sits two points higher on roughly an eighth of the votes, and two points on this board is noise. For clips that should feel casually shot on a phone, Seedance 2.0 is still my default. I'll switch when 2.5 shows a gap bigger than its own error bars.

The usual names, further down

Veo 3.1 with audio has 1398, exactly its June score, and it's now 13th instead of 6th. The 1080p audio variant follows at 1390, then Veo 3.1 Fast with audio at 1385. Kling v3 Pro went from 12th to 19th, with 1354 on 206,221 votes. HappyHorse 1.0 slipped from fourth to tenth while losing only four points. Wan 2.7 is 11th at 1428.

The pattern repeats all the way down. Established models barely budge in score and slide down anyway as new releases land on top. In three months four new releases pushed past Seedance without it losing a single point, and that's not what I expected when I went through this same board in June.

If one of those is on your list anyway, here's where I've written them up at length: Veo 3.1 has its own review, and Kling 3.0 has one too.

Does the second board say the same thing?

Close enough. On Artificial Analysis, with audio, the order runs MiniMax H3 Max (1195), MiniMax H3 (1181), Gemini Omni Flash (1178), then Seedance 2.0 at 1176 and HiDream-O1-Video at 1175, with Wan 3.0 sixth. Grok Imagine 1.5 is eighth, Veo 3.1 eleventh. Those numbers sit on a different scale, so compare the order and ignore the raw values. Two sets of voters with different prompt pools put the same handful of families at the top, and that overlap is the part I actually believe.

What a product still needs that voters don't score

An arena vote rewards the nicer-looking clip. When I animate a product photo I care about something else entirely, which is how much of the original survives: the exact shade of a fabric, the grain of a surface, a logo that has to stay a logo. No public leaderboard measures that, so for me the top ten is a shortlist and the final call comes from running those candidates on the actual image.

That's exactly why I treat the top ten as a shortlist, not a verdict: it points me to a handful of models worth testing, not to the one that will survive a real listing. I wrote up my whole workflow for animating product photos in this guide.

If the pattern holds, Seedance 2.0 should still read 1475 when the December update lands. Most of the models on this board, H3 aside, already sit side by side in Flami, which is where I'll actually find out which one keeps a color true.

Questions about the September board

Which model leads the arena.ai image-to-video board?

minimax-h3 from MiniMax, with 1495 on 57,112 votes as of the September 21, 2026 update. Google's gemini-omni-1.1-flash is second at 1488, but on only 3,734 votes.

Did Seedance 2.0 get worse?

No. Its score is 1475, the same as in June, on 121,427 votes. It's fifth because MiniMax H3, Wan 3.0, Seedance 2.5 and a newer Gemini model arrived above it.

Where does Grok Imagine rank now?

grok-imagine-video-1.5-720p is eighth with 1456. The older 720p version is 12th with 1415 and holds the most votes on the board, 554,050.

Can I run MiniMax H3 on my own hardware in the US?

Not under the current license. Its community license excludes the US from the territory where the weights can be used, along with the EU, the UK and South Korea.

Why don't arena.ai and Artificial Analysis match?

Each board has its own voters and prompts, plus a different rating scale. On September 30, 2026, both put MiniMax H3 first, and both have Seedance 2.0 and Gemini Omni Flash inside their top five.

Checked on September 30, 2026. Sources: the arena.ai update of September 21, 2026, plus the live Artificial Analysis board.

LMArena image-to-video leaderboard: MiniMax H3 moves in on top