Image to Video AI: Animate Any Photo (Full Guide)

Image to video AI is the most controllable way to generate video, because you stop gambling on what the scene looks like: your photo is the first frame, guaranteed, and the model's only job is deciding how it moves. In ClipCraft's video generator that flow is one upload and one sentence: drop the photo in the start-frame slot, describe the motion, and the price of the exact clip you configured sits on the generate button before you spend a token.

The one prompting rule that changes

With a start image loaded, your words should describe motion, not content. The photo already settles what everything looks like; the prompt's job is "slow cinematic push-in, gentle camera drift, natural motion in the scene," which is the actual 72-character prompt in the screenshot below, and nothing more. Describing the scene again mostly gives the model excuses to redraw things you liked. Say what the camera does, say what should move, stop typing.

Which models take a photo, and their quirks

• The Seedance 2.0 family is the image-to-video workhorse: start frames, optional end frames, any clip length from 2 to 15 seconds, sound included, and the character references our consistency workflow leans on.
• Kling 2.5 Turbo Pro takes both a start and an end image, and one quirk is actually a feature here: with a start image set it ignores the aspect-ratio setting and matches your photo instead.
• Hailuo 2.3 animates a first frame only, 16:9, silent, cheap, with the strongest raw motion per token on the board.
• Veo 3.1 and Veo 3.1 Fast take start and end frames with native audio, at the premium end of the price list.

(The lineup thinned in September when OpenAI closed Sora 2's API; it never had a usable image intake anyway, so photo-animators lose nothing.)

The two-frame steering trick

The end-frame slot is the underused one. Give the model where to start AND where to land (two photos) and the clip becomes a controlled move between them: a product turning, a face breaking into the smile from your second shot, a wide establishing drifting into the close-up. The UI walks you into it honestly: on Seedance the end slot stays gray until a start frame exists (the model requires it), and an end frame can't combine with character references, which the route rejects before charging rather than after.

Ratios that fit the photo

Landscape and portrait are one tap, and the Custom button opens the longer list with the use cases labeled right on the chips: 1:1 for posts and avatars, 4:3, 3:4 for print and posters, 21:9 ultrawide for cinematic banners, 9:21 for extra-tall verticals. If the photo needs re-framing before any of that, the aspect-ratio resizer workflow is the two-minute pre-step that saves a re-generation.

The loop that extends a shot

Clips are short everywhere, so the veteran move is the frame-chain: generate clip one, grab its final frame at full quality with the frame extractor, feed that frame back in as the next start image, and repeat. Motion continues, identity holds, and the seams land on cuts you control. Assemble the chain in VideoCraft; the AI-video editing guide is the full playbook for that half.

What it costs, honestly

Prices are per second and shown before you commit: the default Seedance 2.0 Fast at 720p runs 250 tokens a second, so the 5-second clip in my screenshot reads 1,250 on the button, and a 480p draft tier costs about half for finding out whether the motion works. Failures refund automatically, finished clips land in a private My-videos library, and you can close the tab mid-render; the clip finishes and files itself anyway. Video generation needs the Creator or Studio plan; the account itself is free and the galleries stay viewable on any plan.

One photo, one sentence about motion, one visible price. That's the whole technique; start free, bring your best still, and animate it.