What is the best image-to-video AI in 2026?
For control over how a picture moves, Kling 4.0 is the strongest documented option as of 2026-10-18, with start and end frames plus up to 10 keyframes[1]; for keeping a character's face across many clips, Seedance 2.5 takes the most reference images (up to 30)[3]. If you need the lowest per-second price or 4K, Veo 3.1 Lite has the cheapest list rate we found on Google's API and Veo 3.1 goes to 4K[10][9].
What makes image-to-video good
Image-to-video means the model starts from your picture instead of inventing the first frame. That is the most reliable way to keep a face, an outfit and a setting the way you want them. In our own work, clips that start from a still painted off one master image stay on-model far better than clips made from text alone: our explainer Penny is 31 separate clips that read as one character. So we judged the options on:
- Frame control: can you set the first frame, the last frame, or frames in between?
- References: how many extra images can you add to hold a face or a second character?
- Real faces: will the model accept a photo of a real person at all?
- Audio: does it generate dialogue and sound with the picture?
- Length and resolution.
- Price and free access.
The options compared (as of 2026-10-18)
| Model | Frame control | Extra references | Real-face photos | Audio | Longest clip | From, per second |
|---|---|---|---|---|---|---|
| Kling 4.0[1] | Start, end and up to 10 keyframes[1] | Omni Reference: up to 15 assets[1] | Not stated | Yes[1] | 30 s[1] | Not verified |
| Seedance 2.5[3] | First-frame image-to-video | Up to 30 images, 10 videos, 10 audio[3] | Not accepted as direct uploads[5] | Yes[3] | 30 s[3] | $0.2312 at 720p (Replicate)[8] |
| Seedance 2.0 Fast[4] | First-frame image-to-video | Up to 9 images[3] | Not accepted as direct uploads[5] | Yes[7] | 15 s[4] | $0.12 at 720p (BytePlus)[6] |
| Veo 3.1 and Fast[9] | Not stated on the page we read | Up to 3 reference images[9] | Not stated | Yes[11] | 8 s, extendable[9] | $0.10 Fast at 720p[10] |
| Veo 3.1 Lite[9] | Not stated | None[9] | Not stated | Yes[11] | 8 s[9] | $0.05 at 720p[10] |
| Midjourney video[12] | Your image is the first frame[12] | Not stated | Not stated | Not stated | 21 s with extensions[12] | Plan-based, from $10 a month[13] |
| MiniMax-H3[14] | Not stated | Omni Reference[15] | Not stated | Not stated | Not stated | $0.08 at 768P[14] |
| Wan2.2 I2V (open weights)[16] | Image-to-video at 480P or 720P[16] | Not stated | Your own hardware, your own rules | Not stated | Not stated | Free weights, Apache 2.0[16] |
Which keeps the picture's face?
Three things decide it. First, how many views of the face the model can see: Seedance 2.5 accepts up to 30 images and Kling 4.0 up to 15 combined assets, including several views of a character for consistency[3][1]. Second, whether you control later frames: Kling 4.0's keyframes let you pin poses mid-clip[1]. Third, the prompt: in our Seedance renders, describing every extra character's hair, skin and expression in the prompt stopped faces averaging out in group shots. Our character consistency guide has the full method.
If your picture is a photo of a real person, expect friction. BytePlus says Seedance 2.5 and 2.0 do not accept directly uploaded reference images or videos containing real human faces, and offers verification and authorization by that person as one route[5]. CapCut restricted real-face Seedance generations at launch[17], and Character.AI blocks AvatarFX videos made from photos of minors, high-profile politicians and other notable figures[18]. Read is it legal to make AI videos of real people before you try.
Which is cheapest, and which is free?
Per second, Veo 3.1 Lite is the cheapest major option we found at $0.05 at 720p, though it takes no reference images[10][9]. Seedance 2.0 Fast is about $0.12 a second at 720p on BytePlus[6], and in our own Atlas Cloud quotes it came to $0.058 a second for our account. For free use, Vidnoz's image-to-video page says it offers one free generation a day[19], Dreamina gives free daily credits for images and videos and offers Seedance[20], and Wan2.2's weights are free if you have a GPU[16]. The wider price picture is on AI video cost per second.
If you just want a photo to talk
That is a different, cheaper job: animate a face over speech, not a whole scene. Hedra's Character 3 does talking and singing video from audio, priced from 2.5 cents a second on its API[21]; Kling AI Avatar 2.0 takes an image, speech audio and an optional prompt[2]; LivePortrait is open source under the MIT License[22]. See how to make a photo talk with AI.
Verdict
- Pick Kling 4.0 to direct the motion of one image precisely, with keyframes and 30-second clips[1].
- Pick Seedance 2.5 to keep one or several characters consistent across many clips, using up to 30 reference images[3]. Draft on Seedance 2.0 Fast to save money.
- Pick Veo 3.1 Fast or Lite for the lowest API price, and Veo 3.1 for 4K output[10][9].
- Pick Midjourney video if you already make your stills in Midjourney and short, lower-resolution clips are enough[12].
- Pick Wan2.2 if you want to run it yourself, free, at up to 720p[16].
- For the prompt that goes with the picture, see image-to-video prompts.
Where Cast fits
- You do not have a picture yet: describe a character in a sentence and Cast draws four Midjourney looks to pick from (15 credits), then animates the one you choose.
- You want that same character in many short vertical clips: the chosen look is sent as the reference for every later video.
Where it does not:
- Animating your own photo or drawing: Cast does not take image uploads, so it is not an image-to-video tool in the usual sense.
- Making a real person's photo move or talk.
- Horizontal video, 4K, or clips over 15 seconds.
People also ask
- What is the best free image-to-video AI?
- For a daily free clip, Vidnoz's image-to-video page says it offers one free generation a day[19], and Dreamina gives free daily credits for images and videos and offers Seedance[20]. See free image-to-video AI.
- Which image-to-video AI has sound?
- Kling 4.0, Seedance 2.0 and 2.5, and Veo 3.1 generate audio with the video[1][7][3][11]. Seedance 2.5 can also take an audio clip as a reference[3].
- Can I animate a photo of a real person?
- Many services restrict it: BytePlus's Seedance does not accept direct uploads of real faces without verification and authorization[5]. Laws and platform rules also apply; see is it legal to make AI videos of real people.
- How do I stop the face changing in image-to-video?
- Give the model several views of the face as references, keep the first frame from your own picture, and describe the face the same way every time. Our consistency guide covers what worked for us.
Sources
Claims about other companies come from their own pages, read on the date shown. Products and prices change; check the link for the current version.
- Kling AI: AI video generator (read )
- Kling AI: Avatar 2.0 user guide (read )
- BytePlus ModelArk: Seedance 2.5 (read )
- BytePlus ModelArk: Seedance 2.0 (read )
- BytePlus ModelArk: Seedance portrait asset guide (read )
- BytePlus ModelArk: model pricing (read )
- ByteDance Seed: Seedance 2.0 launch (read )
- Replicate: bytedance/seedance-2.5 (read )
- Google: Veo on the Gemini API (read )
- Google: Gemini API pricing (read )
- Google Cloud: generative AI pricing (read )
- Midjourney docs: Video (read )
- Midjourney docs: comparing plans (read )
- MiniMax: API pricing overview (read )
- Hailuo AI: home page (read )
- Wan2.2 on GitHub (read )
- CapCut newsroom: Dreamina Seedance 2.0 in CapCut (read )
- Character.AI blog: AvatarFX (read )
- Vidnoz: image to video AI (read )
- Dreamina: home page (read )
- Hedra: models (read )
- LivePortrait: license (GitHub) (read )
Related answers
- Can I turn a picture into a video with AI for free?
- What should an image-to-video prompt say?
- What is the best AI video generator in 2026?
- How do I make a photo talk with AI?
- Can AI animate a drawing of my character?
- How do I keep an AI character consistent across videos?
Part of the Cast answers library (Other tools and models).
Updated