A still that has already been approved goes in; motion comes out. Choose among the video families that accept a first frame, set duration and aspect ratio family by family, and keep the whole job inside the same workspace as your still pipeline.
22 of 24 families accept a frame · 8 expose first and last frame · 12 expose reference images
The vendors draw the same lines. xAI states that sending an image on its own is image-to-video, and that adding a last frame or reference images turns the request into reference-to-video with that image pinned as the first frame. Vidu lists image to video, reference to video and start-end to video as three separate functions. Kling splits first-frame, first/last-frame, end-frame-only and element control into separate capability rows, and MiniMax states that image-to-video and reference-to-video are mutually exclusive. The availability figures on each card below are read from this site’s own video model catalogue.
One uploaded image becomes the first frame of the clip. This is what the term image to video means wherever it is used precisely. Adding a second frame, or a reference image, moves the request into a different task with different limits.
Availability: 22 of the 24 families in the site catalogue; 21 take one image and seedance-2.5 takes two
Two images are supplied and the model interpolates the motion between them. Vendor support varies: Kling 3.0 Turbo does not support this in the official capability map, while Kling 3.0 Omni, MiniMax H3, Seedance 2.5, Veo 3.1 and Wan 2.7 do. Midjourney Video treats the closing frame as optional.
Availability: 8 families expose it as a mode of its own, each taking two images
Here the input images describe a subject or a scene instead of pinning a frame, and the model composes a new shot around them. Wan 3.0 accepts up to ten reference images at 20 MB each, and MiniMax H3 allows nine. The output framing is no longer borrowed from a supplied still.
Availability: 12 families expose it as a separate mode
The task is narrow, and the vendors keep it narrow. You supply one image and the model animates that image. xAI puts the boundary in plain words: sending an image alone is image-to-video, and combining it with a last frame or reference images is reference-to-video with that image as the pinned first frame. Vidu lists image to video, reference to video and start-end to video as three separate functions rather than three labels for one feature. Kling’s capability map gives first-frame, first/last-frame, end-frame-only and element control a row each. MiniMax states outright that image-to-video and reference-to-video are mutually exclusive. Any tutorial that treats the three as the same thing is contradicted by all four.
This site routes image to video across 22 of the 24 video families in its catalogue. The two that do not accept a frame are viduq3 and viduq3-mix, whose modes are all disabled. Twenty-one of the 22 take a single image; seedance-2.5 takes two. Eight families additionally expose a first-and-last-frame mode, and twelve expose reference images as its own mode. Duration is not uniform either: twenty families run 2 to 15 seconds in one-second steps, seedance-2.5 runs 4 to 30 seconds, vidu-q runs 2 to 16, veo-3.1 is fixed at 4, 6 or 8 seconds, and sora-2 is fixed at 4, 8 or 12. Choosing a family means choosing a set of limits, so the limits are listed per family further down this page rather than averaged into one figure.
Image to video is the last station of a still pipeline, not a replacement for it. Generating a base frame, editing it and then animating it inside one workspace avoids exporting assets between separate apps. Video tasks and image tasks settle through the same points balance, so animating a frame does not multiply the number of accounts involved. This page does not host a workbench of its own: its buttons lead to /create/, where the upload, the prompt and the generation actually happen.
Each rule below names the vendor that wrote it. They differ from one another, and none of them is a setting inside this site.
xAI states that sending an image alone is image-to-video, and that combining it with a last frame or reference images makes it reference-to-video with that image as the pinned first frame. Decide the mode before assembling inputs, because the limits change with it.
MiniMax H3 allows one first frame, one last frame and up to nine reference images, and its documentation says image-to-video and reference-to-video are mutually exclusive. Vidu’s image-to-video endpoint accepts one image only. Wan 3.0 takes one first frame and up to ten reference images.
Vidu allows 50 MB per image and 20 MB per request body. MiniMax allows 30 MB per image and 64 MB per body, and recommends a public URL over base64 for large files. Veo 3.1 caps an input image at 20 MB, and Wan 3.0 caps a reference image at 20 MB as well.
Kling needs both sides at 300 px or more with a ratio between 1:2.5 and 2.5:1. MiniMax takes sides from 256 to 5760 px with a width-to-height ratio between 0.4 and 2.5. Vidu requires the aspect ratio to be narrower than 1:4 or 4:1.
Vidu takes png, jpeg, jpg and webp. Kling takes jpg, jpeg and png. MiniMax goes further and also takes HEIC and HEIF. A frame outside a vendor’s list is rejected before generation begins, so the format is worth checking first.
Vidu caps an image-to-video prompt at 5,000 characters. Kling accepts multi-shot prompts written as shot 1, 2, words; shot 2, 3, words; where each shot runs at least one second, the shot durations add up to the total clip length and each prompt stays within 512 words.
xAI defaults image-to-video output to the input image’s aspect ratio, and an explicit ratio setting overrides that by stretching the frame. Wan 3.0 recommends an adaptive ratio for first-frame work so the output follows the source frame instead of a fixed frame.
Kling documents that generated results are cleared after 30 days. Downloading a clip you intend to keep is part of the workflow rather than an optional extra step at the end.
This page is written against vendor documentation and the site model catalogue. Where the two differ, both are shown rather than merged into one confident sentence.
Read the termsAn image-to-video model treats the frame you upload as the visual anchor and then synthesises change over time: camera movement, fabric and hair motion, a liquid pour, cloud drift. It is a different task from text to video, which has to invent an opening frame first. xAI documents the relationship between the two directly, noting that its text-to-video path generates a first frame and then animates it, so both modes share one image-to-video pipeline underneath. The practical consequence is that the still supplies what exists, and the prompt only decides what moves.
Because the anchor is a real frame, the model never has to guess at the subject. That makes image to video the right mode when a still has already been signed off: a packshot, a portrait, a piece of key art, a diagram, a product label. Reference to video answers a different question, namely what a described subject or scene could look like as a new shot, and start-end to video answers a third, namely how to get from one frame to another. Vendors expose these as separate modes because they accept different inputs and carry different limits, and this site mirrors that split rather than collapsing it.
Model vendors also attach conditions to the frame itself, and those conditions are worth reading before an upload rather than after a rejection. Vidu’s image-to-video endpoint accepts one image, in png, jpeg, jpg or webp, with an aspect ratio narrower than 1:4 or 4:1 and a single file no larger than 50 MB. Kling requires both sides of the frame to be at least 300 px and the aspect ratio to sit between 1:2.5 and 2.5:1. MiniMax takes JPG, JPEG, PNG, WEBP, HEIC and HEIF up to 30 MB, with pixel sides between 256 and 5760 and a width-to-height ratio between 0.4 and 2.5. Veo 3.1 caps an input image at 20 MB. Every rule in this paragraph comes from the vendor that wrote it, not from a setting on this site.
Three steps from a still frame to a finished clip.
png, jpeg, jpg or webp covers most families, and the widest single-image allowance in the vendor documentation is 50 MB. If you plan to use a Kling family, check first that both sides are at least 300 px and that the ratio sits between 1:2.5 and 2.5:1.
Write what should move — a slow pan, drifting clouds, fabric movement — and keep it to what a single shot can show. Duration and aspect ratio are chosen here and they differ by family: 2 to 15 seconds on twenty of them, 4 to 30 on seedance-2.5, five ratios on eighteen of them.
Clips come back as downloadable files rather than permanent storage; Kling documents that result URLs are cleared after 30 days. Save what you intend to keep. If the frame itself needs correction, fix it with image-to-image first and animate the corrected version instead of regenerating.
Every figure below is read from the site model catalogue, family by family, rather than summarised from a marketing page.
Twenty-two video families take a first frame. viduq3 and viduq3-mix do not — all four of their modes are disabled, so they cannot be used for image to video at all.
Twenty-one families take a single image, and seedance-2.5 takes two. Working from a larger reference set is a different mode exposed by twelve families, with its own limits in the vendor documentation.
Twenty families run 2 to 15 seconds in one-second steps. seedance-2.5 runs 4 to 30 seconds, vidu-q runs 2 to 16 seconds, veo-3.1 offers 4, 6 or 8 seconds only, and sora-2 offers 4, 8 or 12 seconds only.
Seventeen families offer 720p and 1080p. Three of them can also drop to 480p and one adds 540p, while sora-2 stops at 720p. Only kling-o3 and veo-3.1 include a 4K option.
Eighteen families offer five ratios — 16:9, 9:16, 1:1, 4:3 and 3:4 — with 16:9 at 720p as the default. Three families add 21:9, and veo-3.1 uses auto, 16:9 and 9:16 instead of that list.
Video tasks and image tasks return the same points unit from the pricing layer, so an animated frame and a generated still settle against one balance instead of two subscriptions. An audio toggle is exposed on four families and a sound-effects toggle on one.
Product and packaging stills are the clearest case. A packshot that has already cleared brand and legal review can go through motion without a reshoot, because the model starts from that exact frame instead of inventing a new scene. Portraits and lifestyle frames follow the same logic, with the subject fixed and only the surrounding movement new. Property and hospitality listings take the same route when on-site video was never shot, and the walk-through becomes a short clip built from the stills that already exist.
Illustrated key art, interface screenshots and diagrams work as well: a screenshot can carry a slow pan for a landing page, and a diagram can gain directional movement in a slide. In each case the frame is the source of truth and the clip is the derivative, so the order matters. If the frame is not right yet, correct it with image-to-image or regenerate it before animating, because the animation cannot be rewound once it has been generated.
Open Create, upload one still, describe the motion and choose the family whose duration and aspect ratio fit your channel.
Open the Create workspace