Which AI video model do people actually prefer?
See the crowd’s votes
Kling 4.0[4]Kuaishou30 s
Seedance 2.5[12]ByteDance (Seed)30 s
Wan3.0[16]Alibaba (Tongyi Lab)30 s- Ray3.2[9]Luma AI20 s
- Vidu Q3[15]ShengShu Technology16 s
- Kling VIDEO 3.0[5]Kuaishou15 s
- Seedance 2.0[11]ByteDance (Seed)15 s
MiniMax H3[7]MiniMax15 s- PixVerse V6[8]PixVerse15 s
Grok Imagine Video 1.5[2]xAI15 s
Gemini Omni Flash[1]Google10 s- HunyuanVideo-1.5[3]Tencent10 s
- Veo 3.1[13]Google DeepMind8 s
Not stated: Veo 3.1 Lite[14], Runway Gen-4.5[10], LTX-2.5[6] — the vendor page publishes no maximum, so no bar is drawn and none is inferred.
| Model | Vendor | Licence | Release | Max resolution | Max duration | Native audio |
|---|---|---|---|---|---|---|
| Proprietary | 2026-08 | 4K | 10 s | Yes | ||
| xAI | Proprietary | 2026-06-16 | 1080p | 15 s | Yes | |
| Tencent | Open weights | 2025-11-20 | 1080p | 10 s | No | |
| Kuaishou | Proprietary | generally available (feature page live); early access rolled out from 2026-09-28 | 4K | 30 s | Yes | |
| Kuaishou | Proprietary | 2026-02-06 | 4K | 15 s | Yes | |
| Lightricks (LTX) | Open weights | Not stated | 3840x2176 | Not stated | Yes | |
| MiniMax | Open weights | 2026-07-31 | 2K | 15 s | Yes | |
| PixVerse | Proprietary | 2026-03 | 1080p | 15 s | Yes | |
| Luma AI | Proprietary | 2026-06-09 | 1080p | 20 s | Not stated | |
| Runway | Proprietary | 2025-12-01 | Not stated | Not stated | Not stated | |
| ByteDance (Seed) | Proprietary | 2026-02-12 | Not stated | 15 s | Yes | |
| ByteDance (Seed) | Proprietary | 2026-07-31 | Not stated | 30 s | Yes | |
| Google DeepMind | Proprietary | 2025-10-15 | 4K | 8 s | Yes | |
| Google DeepMind | Proprietary | 2026-04 | Not stated | Not stated | Yes | |
| ShengShu Technology | Proprietary | Not stated | 1080p | 16 s | Yes | |
| Alibaba (Tongyi Lab) | Proprietary | 2026-08-24 | 1080P | 30 s | Yes |
Sorted alphabetically — this is a catalogue, not a ranking, and the arena has no votes. The licence column is abbreviated; each model page states the full licence, including the community licence name and download location for open-weights models. Maximum resolution and duration are the ceilings the vendor advertises, not defaults. “Not stated” means the vendor page does not publish the value.
Catalogue at a glance
Three ways the 16-model catalogue breaks down. The text carries the counts, recomputed from the vendor-sourced table above; each bar only draws their shares.
Proprietary 13 · Open weights 3
Open weights: MiniMax H3[7], HunyuanVideo-1.5[3], LTX-2.5[6]. The rest are proprietary (API/app only).
Yes 13 · No 1 · Not stated 2
No native audio: HunyuanVideo-1.5[3]. Not stated: Runway Gen-4.5[10], Ray3.2[9].
Text-to-video 16 · Image-to-video 16 · Video edit 12
Every catalogued model publishes text-to-video and image-to-video. No video-edit mode: Kling VIDEO 3.0[5], HunyuanVideo-1.5[3], Vidu Q3[15], PixVerse V6[8].
What the Artificial Analysis image-to-video board shows
Artificial Analysis publishes its own image-to-video leaderboard, built from blind comparisons by a recruited human panel: voters see two clips animated from the same source image and pick the one they prefer. When we checked it on 6 October 2026, MiniMax H3 Max was in first place, followed by MiniMax H3, Gemini Omni Flash, Dreamina Seedance 2.0 720p and HiDream-O1-Video-1.0.[17]
Those positions are theirs, not ours, and so are the names: they are reproduced exactly as Artificial Analysis lists them on that board. Not all of them are publicly released models covered by our catalogue — a board entry is not evidence that a model is generally available, and where our own vendor check could not confirm availability the model is not in the table above. Our arena has no real votes yet, and that table is a vendor-sourced catalogue rather than a ranking. We don't copy their scores, intervals, sample counts or full table here; the complete board, with every model it covers, is on their site.
Source: Artificial Analysis (artificialanalysis.ai), image-to-video leaderboard, checked 2026-10-06. This summary was written by AI Video Arena; it is not reviewed or endorsed by Artificial Analysis.
See the full image-to-video ranking on Artificial Analysis →
How to read this catalogue
Each row is one model, sorted alphabetically — not by quality, because no quality signal exists on this site yet. The columns are the facts a vendor publishes about its own model: who makes it, whether the weights are downloadable or the model is API-only, when it was released, the maximum resolution and clip length the vendor states, and whether the model generates its own audio. Every row carries a numbered marker pointing at the vendor page the row was read from, listed in full at the bottom of the page. Where a vendor does not state a value we write Not stated rather than guess.
Two habits make the table more useful. First, read the licence column before anything else: open weights you can download and run locally and a hosted API are different purchasing decisions, and the arena will never collapse that difference into a single number. Second, treat maximum resolution and maximum duration as ceilings, not defaults — they are the best case the vendor advertises, usually at the highest tier, and say nothing about what a given prompt will actually produce. The licence checkboxes below the tabs are progressive enhancement: every row is in the HTML without JavaScript, and the filter only hides rows in your browser.
Featured battles
Comparable models placed side by side on the facts their vendors publish. No win shares are shown anywhere, because nobody has voted yet.
Generate your own clips
The arena only works if voters have hands-on context. AIContentDrop is a credit-based multi-model AI video and image platform where you can run current models yourself before you judge other people's outputs. See the plan and credit details or drive generations from your own pipeline via the REST API.
Prefer staying inside your editor? Wire the platform into your tools through MCP.
The crowd has voted 425,931 times
Not on this site — on Design Arena's. These are their blind pairwise preference votes: visitors there pick between two unlabelled clips, every model's Elo starts at 1200 and is fitted with the standard Bradley-Terry model, and the margin shown is plus-or-minus one standard error from that fit. We checked their board on 2026-10-21[18]. Our own arena here still has zero votes until voting starts; the table below is theirs, quoted in full, so you can see what a crowd-voted video board looks like while ours waits for its first ballot.
- 1Gemini Omni 1.1 FlashGoogle1363
- 2Gemini Omni FlashGoogle1346
- 3
HeyGen Video 1.0 (1080p, Ultraquality)HeyGen1337 - 4Wan 3.0Alibaba1329
- 5P-Video-2-Pro (Cost)Pruna AI1313
- 6FLUX 3 MaxFal1298
- 7FLUX 3 VideoBlack Forest Labs1289
- 8MiniMax H3 MaxFal1281
- 9MiniMax H3MiniMax1280
- 10HeyGen Video 1.0HeyGen1273
| Rank | Model | Vendor | Elo | Win rate | Votes |
|---|---|---|---|---|---|
| 1 | Gemini Omni 1.1 Flash | 1363 | 68.3% ± 1.5% | 3,896 | |
| 2 | Gemini Omni Flash | 1346 | 64.3% ± 1.0% | 8,497 | |
| 3 | HeyGen Video 1.0 (1080p, Ultraquality) | HeyGen | 1337 | 67.4% ± 3.1% | 856 |
| 4 | Wan 3.0 | Alibaba | 1329 | 63.5% ± 1.5% | 3,891 |
| 5 | P-Video-2-Pro (Cost) | Pruna AI | 1313 | 63.6% ± 2.4% | 1,484 |
| 6 | FLUX 3 Max | Fal | 1298 | 62.1% ± 1.7% | 3,106 |
| 7 | FLUX 3 Video | Black Forest Labs | 1289 | 64.5% ± 1.1% | 7,497 |
| 8 | MiniMax H3 Max | Fal | 1281 | 53.2% ± 0.7% | 20,029 |
| 9 | MiniMax H3 | MiniMax | 1280 | 63.7% ± 1.2% | 5,850 |
| 10 | HeyGen Video 1.0 | HeyGen | 1273 | 57.4% ± 2.8% | 1,160 |
| 11 | Seedance 2.5 | ByteDance | 1259 | 57.4% ± 1.7% | 3,433 |
| 12 | Seedance 2.0 | ByteDance | 1255 | 58.0% ± 0.9% | 11,465 |
| 13 | Seedance 2.0 Mini | ByteDance | 1255 | 60.2% ± 0.8% | 13,880 |
| 14 | Seedance 2.0 Fast | ByteDance | 1253 | 57.6% ± 0.9% | 11,701 |
| 15 | Happy Horse 1.1 | Alibaba | 1227 | 49.5% ± 1.5% | 4,267 |
| 16 | Grok Imagine | xAI | 1222 | 67.6% ± 0.4% | 51,441 |
| 17 | Happy Horse 1.0 | Alibaba | 1212 | 64.0% ± 0.8% | 15,073 |
| 18 | Veo 3 Fast | 1184 | 63.0% ± 0.9% | 10,922 | |
| 19 | Veo 3.1 Lite | 1177 | 57.5% ± 0.6% | 27,782 | |
| 20 | Veo 3.1 Fast | 1171 | 56.8% ± 0.5% | 36,032 | |
| 21 | Veo 3.1 | 1169 | 59.7% ± 0.7% | 18,087 | |
| 22 | Wan 2.7 | Alibaba | 1167 | 54.6% ± 0.8% | 15,498 |
| 23 | Kling v3 Pro | KlingAI | 1167 | 61.5% ± 2.2% | 1,922 |
| 24 | LTX 2.5 Fast | Lightricks | 1166 | 43.9% ± 1.4% | 4,866 |
| 25 | LTX 2.5 Pro | Lightricks | 1161 | 43.5% ± 1.4% | 4,873 |
| 26 | Veo 3 | 1153 | 58.1% ± 3.3% | 834 | |
| 27 | P-Video 2 | Pruna AI | 1149 | 42.2% ± 2.4% | 1,563 |
| 28 | Sora 2 | OpenAI | 1142 | 58.0% ± 0.4% | 52,022 |
| 29 | Sora 2 Pro | OpenAI | 1141 | 57.4% ± 1.9% | 2,654 |
| 30 | Kling O3 (Omni) | KlingAI | 1130 | 36.2% ± 1.3% | 5,287 |
| 31 | P-Video 2 Draft | Pruna AI | 1118 | 36.7% ± 2.4% | 1,498 |
| 32 | Seedance 1.5 Pro | ByteDance | 1102 | 52.5% ± 0.5% | 47,216 |
| 33 | Ray 3.2 | Luma Labs | 1096 | 42.0% ± 2.2% | 1,996 |
| 34 | Kling v2.6 Pro | KlingAI | 1093 | 51.0% ± 0.4% | 47,447 |
| 35 | Ray 3.14 | Luma Labs | 1092 | 54.4% ± 2.0% | 2,444 |
| 36 | Kling v2.5 Turbo Pro | KlingAI | 1077 | 49.2% ± 0.4% | 58,632 |
| 37 | MiniMax Hailuo-2.3 (Pro) | MiniMax | 1076 | 48.7% ± 0.5% | 45,762 |
| 38 | MiniMax Hailuo-2.3 (Standard) | MiniMax | 1069 | 48.0% ± 0.4% | 48,383 |
| 39 | LTX 2.3 | Lightricks | 1059 | 34.9% ± 2.0% | 2,110 |
| 40 | Glam.ai 1.0 | Glam.ai | 1049 | 44.5% ± 0.5% | 32,356 |
| 41 | P-Video (1080p) | Pruna AI | 1040 | 43.6% ± 1.7% | 3,469 |
| 42 | Seedance 1.0 Pro Fast | ByteDance | 1020 | 41.0% ± 0.4% | 64,552 |
| 43 | P-Video (720p) | Pruna AI | 1017 | 41.9% ± 1.5% | 4,326 |
| 44 | Kandinsky 5.0 Pro | AI Forever | 993 | 36.6% ± 0.5% | 36,880 |
| 45 | Kandinsky 5.0 | AI Forever | 965 | 33.1% ± 0.4% | 55,492 |
| 46 | LTX-2 19B | Lightricks | 946 | 29.4% ± 0.4% | 48,909 |
| 47 | Ray 2 | Luma Labs | 938 | 34.4% ± 1.9% | 2,306 |
| 48 | Ray 2 Flash | Luma Labs | 924 | 31.1% ± 1.7% | 2,849 |
Source: Design Arena (designarena.ai/leaderboard/video), checked 2026-10-21. Updated Oct 21, 2026, 11:07 PM GMT+3 by the publisher.
Three external sources behind this page
Everything on this site that comes from a third party comes from one of these three. Each card says what the source covers and jumps to where we use it; none of their figures, layouts or images is republished here as ours.
Artificial Analysis
An independent image-to-video leaderboard built from blind comparisons by a recruited human panel. Their terms allow a named, attributed prose summary only, so ours names their leading models without republishing their table or scores.
Design Arena
A public crowd-voted video leaderboard — blind pairwise preference votes fitted to an Elo scale by Design Arena, not by us. Their board is quoted in full, marked as theirs, in the crowd-votes section.[18]
Source: Design Arena (designarena.ai/leaderboard/video), checked 2026-10-21.
OpenRouter
The OpenRouter API routes image-generation models and publishes their prices, context windows and providers. That is image-generation reference data, not video-arena evidence — it says nothing about video models.[19]
Source: OpenRouter compare view, checked 2026-10-22.
Sources
- Google — Gemini Omni Flash, vendor documentation https://ai.google.dev/gemini-api/docs/models/gemini-omni-flash — checked 2026-10-06
- xAI — Grok Imagine Video 1.5, vendor documentation https://docs.x.ai/developers/model-capabilities/video/generation — checked 2026-10-23
- Tencent — HunyuanVideo-1.5, vendor documentation https://huggingface.co/tencent/HunyuanVideo-1.5 — checked 2026-10-06
- Kuaishou — Kling 4.0, vendor documentation https://kling.ai/feature/kling-4-ai-video-generation — checked 2026-10-23
- Kuaishou — Kling VIDEO 3.0, vendor documentation https://kling.ai/quickstart/klingai-video-3-model-user-guide — checked 2026-10-06
- Lightricks (LTX) — LTX-2.5, vendor documentation https://huggingface.co/Lightricks/LTX-2.5 — checked 2026-10-06
- MiniMax — MiniMax H3, vendor documentation https://www.minimax.io/blog/minimax-h3 — checked 2026-10-06
- PixVerse — PixVerse V6, vendor documentation https://pixverse.blog/en/updates/pixverse-launches-v6-advancing-ai-video-generation/ — checked 2026-10-06
- Luma AI — Ray3.2, vendor documentation https://lumalabs.ai/news/introducing-ray-3-2 — checked 2026-10-06
- Runway — Runway Gen-4.5, vendor documentation https://runway.com/research/introducing-runway-gen-4.5 — checked 2026-10-06
- ByteDance (Seed) — Seedance 2.0, vendor documentation https://seed.bytedance.com/en/seedance2_0 — checked 2026-10-06
- ByteDance (Seed) — Seedance 2.5, vendor documentation https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5 — checked 2026-10-06
- Google DeepMind — Veo 3.1, vendor documentation https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/ — checked 2026-10-06
- Google DeepMind — Veo 3.1 Lite, vendor documentation https://ai.google.dev/gemini-api/docs/models — checked 2026-10-06
- ShengShu Technology — Vidu Q3, vendor documentation https://www.vidu.com/vidu-q3 — checked 2026-10-06
- Alibaba (Tongyi Lab) — Wan3.0, vendor documentation https://www.alibabacloud.com/en/news/product/video-generation-model-wan30-officially-launched-nsb — checked 2026-10-06
- Artificial Analysis — image-to-video leaderboard https://artificialanalysis.ai/video/leaderboard/image-to-video — checked 2026-10-06
- Design Arena — video leaderboard https://www.designarena.ai/leaderboard/video — checked 2026-10-21
- OpenRouter — image model compare view https://openrouter.ai/compare/google/gemini-3.1-flash-image/meta/muse-image/openai/gpt-image-2.5-sunburst — checked 2026-10-22
Each [n] above is the vendor page the fact beside the marker was read from, on the date shown. Where a vendor states nothing, the table says “Not stated” and no value is inferred from anywhere else.