Workflow
From a still to a video: turning a picture you already have into a clip
A still you already have is the cheapest thing you can bring to a video model: the composition is settled, so the whole render goes into motion rather than into inventing a scene. The order of operations is to pick the kind of model that matches what your picture is for, crop and upload it so it is neither rejected nor silently reframed, then settle the shot on a short draft before paying for a long one.
Decide whether your picture is a frame or a description
An image-to-video model treats your picture as the opening frame. The render begins on those exact pixels, so the picture settles the content and the shape of every frame after it. A reference-to-video model does something else with the same file: your pictures are evidence about who or what appears, none of them becomes a frame, and composing the shot stays the model's job.
The catalog prints the capability on every card, so filter to video and read the label rather than guessing from a name. That capability belongs to the model and not to the run, so one that wants an opening frame refuses a request without one, and one that reads no pictures refuses a request that carries them; keeping a character consistent covers why being refused is the behaviour you want here.
Which one you need depends on how much you have already decided. If you generated and framed the still yourself and you know the moment the clip opens on, hand it over as the frame. If a character has to appear but the shot still needs composing, hand it over as evidence and write the staging.
Crop and check the rotation before you upload
Because the still becomes the first frame, its shape becomes the clip's shape. Models that open on a picture do not offer an aspect ratio at all: the field is absent from the form rather than disabled, and sending one over the API comes back as an error naming it. So crop in your image editor while the file is still an image. Cropping the finished clip means paying to render pixels you throw away, and re-framing a composition that was built for the wider box.
Check the rotation while you are in there. Photographs off a phone carry an EXIF Orientation tag and your browser quietly applies it: the initial value of the CSS image-orientation property is from-image, which MDN describes as meaning that "the EXIF information contained in the image is used to rotate the image appropriately". A preview can therefore sit upright while the stored pixels lie on their side. If a clip comes back rotated, re-export the still with the rotation baked into the pixels.
Everything else riding on that frame comes with it: a watermark, a letterbox bar, a border, a colour cast you meant to fix later. The opening frame is not a draft the model tidies up on the way past.
Get the file to a URL the model can fetch
Generation endpoints take pictures as addresses rather than as bytes, because the provider performing the render fetches them itself. That is also why a link to one of your own generations does not work: those are served behind your credentials and the renderer has no session to use. Upload the file and attach the address that comes back, which carries an unguessable id and is reachable without signing in.
You do not have to host anything. The reference box on a model's page uploads the file and writes the address into the field for you, and the API reference documents the same endpoint for scripts, where you post raw bytes rather than a multipart form.
Rejections here come from a short list, and the error names the field. The Content-Type header is required and has to match the body, because the server reads the file's magic bytes and renaming a screenshot does not change what it is. The accepted formats are the ordinary ones: JPEG, PNG, WebP, BMP and GIF. Uploading needs a positive credit balance. An address you supply yourself has to be https on a public host, so loopback and private ranges are turned down. Uploads are held for the same window as generated output, after which the address answers as expired rather than as missing. Calling the API from your code covers the request side.
One take, so the prompt is the edit
There is no timeline. The model renders the whole clip in one pass, and there is no point mid-render where you step in, cut a beat or change your mind about the camera. Since the picture already answered what the scene looks like, spend the words on what changes: how the subject moves, how the camera moves, and where each of them ends up.
Name the camera every time. Leaving it out does not hold the frame still, it hands the choice to the model, and a drift you did not ask for looks exactly like one you did. Writing prompts without a filter goes further into wording.
Budget one beat rather than a scene. Cutting, DeLong and Brunick, analysing 295 English-language feature films in Cognitive Research: Principles and Implications (2018), report that "shots have generally gotten shorter, decreasing more or less linearly from a mean of about 10 s/shot in 1950 to about 4 s/shot in 2010". A take in commercial film carries roughly one action, and your clip is one take, so an action written for a whole scene arrives hurried whatever length you buy.
Settle the shot on a short draft
Length and resolution compound, so the order that wastes least is a short draft at a modest resolution until the motion reads, then one long run at the setting you want. Choosing resolution and length has that trade in full, and every model page does the arithmetic live from the catalog.
Keep the still fixed between takes so the sentence is the only thing that moved. A run that fails at the provider refunds itself; a run that succeeded in a shape you did not want does not, which is the whole argument for cropping first. Download what you decide to keep. See keeping a character consistent when the second clip has to match the first, and storyboards and previz when there are ten of them.
Frequently asked questions
Can I set an aspect ratio when the clip starts from my picture?
No. The still is the first frame, so its shape is the clip's shape. Models that open on a picture expose no ratio, and sending one to the API is refused rather than ignored, because a clip returned in a shape you did not choose, with no error explaining it, would be worse. Models that work from a description alone do take a ratio, and on some it is required.
Why was my reference image rejected?
The error names the field, and the causes are few: the model reads no reference pictures, or wants a different number than you sent; the address is not https, or points at a private host; the Content-Type header does not match the file's bytes; the format is not an accepted one; or the account has no credit balance to upload against.
Should I use an opening frame or references for a recurring character?
Use an opening frame when you know the exact moment the clip should start on, which is common when you generated and framed that still yourself. Use references when the character has to appear but the shot still needs composing: the pictures constrain who is in it, not where the camera is.