AI Video Generators Compared: Veo, Kling, Seedance, Wan, Hailuo and Runway

What this page compares, and what it doesn't

Everything here comes from each vendor's official documentation, checked in November 2026. It isn't a ranking: we didn't generate clips to score quality, we leave out prices (they change often and depend on the plan), and we don't name a winner. What you do get is what each model can do, how it expects to be prompted, and which options it offers.

Sora isn't included because OpenAI has retired it: the API stopped working on 24 September 2026. If you're moving off Sora, see the Sora alternatives guide.

Current versions

• Google Veo 3.1: Veo 3.1, Veo 3.1 Fast and Veo 3.1 Lite (in preview). Google says Veo 3.1 Fast will be retired "on or after" 17 November 2026.
• Kling: Kling 4.0 is current; Kling 3.0 and 3.0 Omni are still offered.
• Seedance (ByteDance): Seedance 2.0 (standard, fast and mini) and Seedance 2.5.
• Wan (Alibaba): Wan 3.0 is the newest, API only; Wan 2.7, 2.6 and 2.5 remain on the API, and Wan 2.2 is the latest open-weights release.
• MiniMax Hailuo: MiniMax-H3 ("Hailuo 3.0") is current; Hailuo 2.3, Hailuo-02 and T2V-01-Director remain on the API.
• Runway: Gen-4.5 for generation and Aleph 2.0 for video editing.

Modes and clip length

• Veo 3.1: text-to-video, image-to-video, first and last frame, reference images (not in Lite) and extension. Clips of 4, 6 or 8 s (8 s only with references); each extension adds 7 s, up to 37 s in total. 16:9 or 9:16; 720p or 1080p, with 4K on the standard model only.
• Kling 4.0: two input modes that can't be combined, Omni Reference (default, up to 15 assets) and Keyframes (up to 10 keyframes). Editing of up to 5 clips and multi-shot. 3 to 30 s (10 s by default); 720p, 1080p or 4K; 21:9, 16:9, 1:1, 9:16 or auto.
• Kling 3.0: text-to-video, image-to-video, start and end frames, Elements (a character built from 2–4 images or a video) and up to 6 shots. 3 to 15 s; 720p or 1080p.
• Seedance 2.0: text-to-video, first frame, first and last frame, omni reference (up to 9 images, 3 videos and 3 audio files), editing and extension. 4 to 15 s; 480p and 720p (1080p and 4K on standard only); 21:9 through 9:16. It doesn't directly accept reference images or videos with real human faces.
• Seedance 2.5: 4 to 30 s, up to 1080p at 10-bit and up to 50 references; accepts storyboards of up to 15 panels.
• Wan 3.0: 2 to 30 s, up to 1080p, optional audio and omni reference. Wan 2.7: 2 to 15 s, 720p or 1080p, five aspect ratios. Wan 2.2 (open weights): no sound.
• MiniMax H3: text-to-video, first and/or last frame, reference (up to 9 images, 3 videos and 3 audio files) and editing. 4 to 15 s; 768p or 2K; 24 fps. Hailuo 02 and 2.3: 6 or 10 s at 768p, or 6 s at 1080p.
• Runway Gen-4.5: text-to-video and image-to-video. We couldn't verify durations, aspect ratios or limits in official sources.

Audio and dialogue

• Veo 3.1: native dialogue, sound effects and ambience. Dialogue in quotes with the speaker named, "SFX:" for effects and "Ambient noise:" for background sound.
• Kling 4.0: stereo audio with lip sync in 9 languages (Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi). Dialogue goes in quotes with the language stated and the voice described. Kling 3.0 documents five languages.
• Seedance 2.0: dialogue in { }, sound effects in < >, music in （ ） and one language per line. Seedance 2.5 documents 11 languages.
• Wan: versions 2.5 to 2.7 add matching music or effects if you don't supply audio; Wan 3.0 has optional audio; Wan 2.2 and 2.1 are silent. For speech, quote the line and describe emotion, tone, speed, timbre and accent.
• MiniMax H3: dialogue in 11 languages, using <d> tags and a speaker ID. Audio for Hailuo 02 and 2.3 isn't documented.
• Runway Gen-4.5: press coverage mentions native audio, but we couldn't confirm it in official documentation.

Prompt language and negative prompts

• Veo 3.1: English prompts, per the model card. A separate negativePrompt field takes a list of terms ("text, watermark"), not sentences with "no".
• Kling 4.0: the guide advises against negative prompts because they can increase errors; it only accepts specific exclusions such as no subtitles or no background music.
• Seedance: no negative field. Use constraint phrases inside the prompt, such as "avoid generating any text or subtitles" or "do not generate a watermark".
• Wan: prompt and negative in Chinese or English. A negative_prompt API field (up to 500 characters on 2.7), plus phrases like "No background music." inside the prompt. Note that 2.7 rewrites your prompt with an LLM by default (prompt_extend).
• MiniMax H3: the soundscape section is written in English; negatives aren't documented. MiniMax suggests running prompts through its H3-Context-IR rewriter first.
• Runway Gen-4.5: the guide says negative phrasing isn't supported and may produce unpredictable or even opposite results. Describe only what should happen.

Special syntax by model

• Veo 3.1: cinematography + subject + action + context + style; multi-shot with [00:00-00:02] timestamps; references phrased as "Using the provided images for…".
• Kling 4.0: uploaded assets tagged @Image1, @Video1 or @Element1, a "0–X s" timeline and, for edits, "Edit @Video1 … keep everything else unchanged". Kling 3.0 shots look like "Shot 1 (3s):".
• Seedance: literal symbols （ ） < > { } 【 】, "Shot 1 / Shot 2" with one camera move each, and references as "Image 1", "Video 1". In 2.5, whole-second timings ("0-3s, 3-7s").
• Wan: an overall description followed by "Shot 1 [0–3 s] … Shot 2 [3–6 s] …"; a unique label per character and the action before the dialogue.
• MiniMax H3: integrated_multimodal_description, overall_soundscape and non_diegetic_music sections; dialogue as (S1) plus <d>[English] …</d>; cuts as "At 00:03.500". Hailuo 02 and 2.3: 15 bracketed camera commands.
• Runway Gen-4.5: positive prose only; "locked-off" for a still camera; complex moves described phase by phase.

Where you can use each one

• Veo 3.1: GA models on Vertex AI. On the Gemini API, the Veo 3.1 preview models were switched off on 22 October 2026. We haven't verified access through the Gemini app or Flow. Runway's API also offers Veo 3.1.
• Kling: the kling.ai website and app; Runway's API lists Kling 4.0.
• Seedance: BytePlus ModelArk, Volcengine Ark, Dola and BytePlus Playground, plus Runway's API. Some sources also place it in Dreamina and CapCut.
• Wan: Wan 2.2 open weights (Apache 2.0) on GitHub; Wan 3.0 and 2.5–2.7 via API on Alibaba Model Studio; Runway's API offers Wan 3.0.
• MiniMax Hailuo: H3 weights on Hugging Face under MiniMax's own community licence, plus the MiniMax API and Runway's API. Secondary sources also mention the Hailuo app.
• Runway: Gen-4.5 on paid plans. Its API also acts as a gateway to other vendors' models.

One prompt, rewritten for each model

The idea: a potter in her sixties lifts a freshly glazed bowl from the wheel in her workshop at dusk while it rains, and says one short line. Slow push-in, warm lamp light, documentary style, no on-screen text. Here it is in the form each vendor documents:

Slow dolly-in, medium shot at eye level. A potter in her sixties with a clay-stained linen apron lifts a freshly glazed blue bowl from the wheel in a small workshop at dusk. Warm light from a single desk lamp, rain running down the window behind her. Documentary style, natural colours, soft focus on the background. The potter says quietly, "This one stays with me." SFX: the wheel slowing to a stop. Ambient noise: steady rain on the glass.

text, subtitles, watermark, logo

Scene: a small pottery workshop at dusk, rain running down the window, lit by a single warm desk lamp. Style: realistic, natural documentary look. Camera: medium shot at eye level with a slow push-in throughout.
0–4 s: a potter in her sixties with a clay-stained linen apron stops the wheel and lifts a freshly glazed blue bowl with both hands.
4–10 s: she turns the bowl toward the lamp, studies it and smiles.
Audio: steady rain on the glass and the wheel slowing down. Potter (calm, warm voice of a woman in her sixties, slow pace, English): "This one stays with me." No subtitles, no background music.

A potter in her sixties with grey hair tied back and a clay-stained linen apron lifts a freshly glazed blue bowl from the wheel. Small pottery workshop at dusk, rain on the window. Warm light from a single desk lamp, soft shadows. Shot 1: medium shot, slow push-in as she lifts the bowl. Shot 2: close-up of her hands turning the bowl toward the lamp, camera fixed. Natural documentary style, realistic textures, high detail. The potter says {This one stays with me.} <rain tapping on glass> <the wheel slowing down> Avoid generating any text or subtitles. Do not generate a watermark or logo.

A potter in her sixties lifts a freshly glazed blue bowl from the wheel in a small workshop at dusk, rain on the window, warm desk-lamp light, documentary realism. Shot 1 [0–4 s] medium shot, the camera pushes in slowly as the potter stops the wheel and lifts the bowl. Shot 2 [4–8 s] close-up, fixed camera, the potter turns the bowl toward the lamp, then says "This one stays with me." in a calm, warm, unhurried voice with a soft, slightly husky timbre and a neutral accent. Sound: steady rain on the glass.

subtitles, on-screen text, watermark, background music

integrated_multimodal_description: [Shot 1] A medium shot frames a potter in her sixties lifting a freshly glazed blue bowl from the wheel in a small workshop at dusk, lit by a single warm desk lamp, rain running down the window; the camera pushes in with small amplitude at slow speed. At 00:04.000, the camera cuts to a close-up of the bowl turning in her hands as the potter (S1) says: <d>[English] This one stays with me.</d>
overall_soundscape: Steady rain taps on the window glass. A potter's wheel hums and slows to a stop.
non_diegetic_music: N/A

A potter in her sixties lifts a freshly glazed blue bowl from the wheel in a small workshop at dusk, warm lamp light and rain on the window [Push in], then turns the bowl toward the lamp and smiles [Static shot].

The camera dollies in slowly on a potter in her sixties as she lifts a freshly glazed blue bowl from the wheel in a small workshop at dusk. She turns the bowl toward a warm desk lamp and smiles. Rain runs down the window behind her. Natural documentary look, soft shadows, a clean frame showing only the workshop and the potter.

Notice what changes: negatives move to a separate field in Veo and Wan, become constraint phrases in Seedance and Kling, and vanish in Runway; and the Hailuo 2.3 and Runway versions leave out sound because their audio isn't documented.

Which one to try for…

We won't tell you which is best, but we can tell you which ones document the feature you need:

• Dialogue in languages other than English: Kling 4.0 lists nine (including Spanish, Portuguese, German and French); Seedance 2.5 and MiniMax H3 list 11 each, including Spanish. Check the exact list in each guide.
• Open weights you can run yourself: Wan 2.2 (Apache 2.0, no sound). MiniMax H3 also publishes weights, under its own community licence, which is worth reading.
• Long clips in one generation: Kling 4.0, Seedance 2.5 and Wan 3.0 reach 30 s. Veo 3.1 reaches 37 s by chaining extensions.
• A separate negative prompt field: Veo 3.1 (negativePrompt) and Wan (negative_prompt).
• Lots of character or style references: Seedance 2.5 (up to 50), Kling 4.0 (up to 15 assets), Seedance 2.0 and H3 (up to 9 images, 3 videos and 3 audio files).
• Camera control through fixed commands: Hailuo 02 and 2.3, with 15 bracketed commands.
• First and last frame: Veo 3.1, Kling 3.0, Kling 4.0 (Keyframes), Seedance 2.0 and H3.
• Editing footage you already have: Kling 4.0, Seedance 2.0, H3 and Runway Aleph 2.0.
• Vertical 9:16 video: documented by Veo 3.1, Kling 4.0, Seedance 2.0 and Wan 2.7.

How to run your own comparison

• Use the same scene everywhere, adapted to each model's syntax as in the examples above.
• Match duration, aspect ratio and resolution wherever the model allows it.
• Generate each prompt at least two or three times; one result tells you little.
• Change one thing per attempt (camera, light, dialogue) so you know what made the difference.
• Write down the exact model version and the date: models change and your notes go stale.

To avoid rewriting the scene by hand for every format, PromptNow has one prompt generator per model (Veo, Kling, Seedance, Wan, Hailuo and Runway): fill in the form once and get the prompt in that tool's syntax, with any separate fields it uses.