# Image to video AI, step by step — animate a photo or a Character > Animate a reference photo or a Character in eroq — pick the right engine, use first and last frames on Seedance 2.5, write prompts that describe motion only. Published 2026-09-04 · eroq.ai — canonical: https://eroq.ai/blog/image-to-video-guide Text-to-video asks the engine to invent everything: the subject, the framing, the light, and then the motion on top. Image-to-video takes the first three off its plate. You hand it a still that already looks right and ask for one thing — make it move. It is the most controllable way to get a specific face, product or composition into a clip, and it is the workflow behind most of the "how did they keep that consistent" videos you have seen. Here is how it works in the eroq studio, from source image to finished clip. ## Where the source image comes from The [Video tool](/studio/video) accepts a reference image from three places: - **A photo you upload.** Any still you have the rights to use. Fictional adult subjects only — the studio does not animate real, identifiable people without their consent, and that rule is enforced at the account level. - **A Character.** A saved [Character](/studio/characters) carries reference photos, so picking one as the source hands the engine the face you have already approved. - **A render from your Library.** Every image you generate is saved automatically. The Animate button on an image card sends it to the Video tool as an image-to-video source, with the original recipe one click away under Remix. If you do not have the still yet, make it first in [the Image tool](/studio/image): Image One for photoreal, Image Anime for illustration, Image Art for painterly. Ten credits per image, batch 1 to 4, and you can pin a face with reference photos there too. Generating the still separately is cheaper than re-rolling a video until the framing is right — a 5-second clip costs 100 credits on Motion One, an image costs 10. ## Choosing an engine Image-to-video uses the same engine picker as text-to-video, and the trade-offs are the same: - [Motion One · Wan](/models/eroq-motion-one) is available to every account, including free, and is the uncensored engine. 5 or 10 s, 480p or 720p, 100 or 180 credits. - [Seedance 2.5](/models/seedance-2-5) goes from 4 to 30 s at up to 1080p, honors a seed, and is the engine with first/last-frame interpolation (more below). 180 credits per 5 s, Creator plan and up. - Seedance 1.0 Pro and Lite cover 3 to 12 s with a seed; Lite is the budget option at 60 credits per 5 s on Hobby and up. - Kling 2.5 Turbo is known for fluid motion and faces. Hailuo 02 for cinematic physics and dramatic light. Veo 3 Fast renders a fixed 8 s with native sound. Kling and Veo are Creator and up; Hailuo, Hobby and up. Length is also capped by your plan — 10 s free, 15 s Hobby, 20 s Creator, 30 s on Studio and above. For a portrait or a product turn, 5 s is usually enough; save the long clips for slow camera moves where the still can carry the frame. ## Write the prompt for motion, not for the picture This is the part most people get wrong. The photo already tells the engine who is in the frame, how it is lit and where the camera stands. If your prompt re-describes all of that, you give the engine permission to drift from the still — a different jacket, a warmer light, a face that is almost hers. Describe what moves and nothing else. A complete motion-only prompt for a portrait: > She slowly turns her head toward the window and exhales, a strand of hair lifting in the draft; slow push-in on a soft portrait lens, the curtain behind her swaying once, dust drifting through the shaft of window light, calm and unhurried, nothing else in the frame moves. Notice what is missing: no description of her face, clothes or the room. Three things are present — the subject's motion, the camera's motion, and the tempo. That is the whole recipe. A few rules that follow from it: 1. **One camera move per clip.** Pick it from the camera move chips (Push-in, Orbit, Slow pan and the rest) or write it in the prompt — written directions like "dolly" or "wide shot" are detected either way. 2. **Name the tempo.** Calm, Dreamy, Dynamic, Tense or Chaotic on the Director's rack, or a word in the prompt. Without it, engines tend to over-animate. 3. **Say what stays still.** "Nothing else moves" is a legitimate instruction and a useful one. 4. **Use the negative prompt on engines that support it** for the drift you keep seeing — "extra fingers, warping text, changing hairstyle". 5. **Lock the seed on Seedance** when a take is close: same seed, small prompt change, and you get a variation instead of a reroll. ## First and last frame on Seedance 2.5 Seedance 2.5 accepts two stills instead of one: a first frame and a last frame, and it interpolates the motion between them. That turns image-to-video from "animate this" into "get from here to there", which is a different tool. It is the move for reveals and transformations: a product closed, then open; a room empty, then full; the same character at the door, then at the window. Two things make it work. Keep the frames consistent — same subject, same framing, same light, ideally two images from one batch or two frames from one clip — and describe only the transition in the prompt, in the same motion-only register as above. The [first/last frame](/glossary/first-last-frame) glossary entry has the mechanics; the short version is that mismatched frames give you a morph, and morphs are rarely what you wanted. ## Mistakes worth skipping - **Busy stills.** A source image with six people and a crowd will animate six people and a crowd, badly. Give the engine one subject. - **Describing the past.** "A woman who has just walked in" is a backstory, not a motion. "Walks in from the left and stops" is a motion. - **Fighting the frame.** Asking a portrait to become a wide shot is asking the engine to invent a room. Use Pull-back for a little air, or generate the wide still and animate that. - **Re-rolling instead of re-stilling.** If the framing is wrong, fix the image. It is ten credits. Failed renders refund themselves, and every clip is saved to your Library with its source and recipe, so the loop is cheap to iterate: still, motion prompt, take, adjust. ## FAQ ### Which engine should I start with for image-to-video? Motion One if you are on the free tier or need uncensored output; Seedance 2.5 if you want first/last frames, a seed and up to 30 s; Kling 2.5 Turbo when the clip is mostly a face. ### Can I animate a Character directly? Yes. Pick the Character as the source and the engine animates its reference photo. @mention it in the prompt as well so the persona rides along. ### Does the prompt need to describe the person in the photo? No — and it should not. The still carries the subject; the prompt carries the motion, the camera and the tempo. Have a still that deserves to move? [Open the Video tool](/studio/video) and drop it in, or read the [image-to-video overview](/tools/image-to-video) first.