Choose the role of each input
| Input | What it guides | VidFlux | PromptToVisual |
|---|---|---|---|
| First frame | The starting composition | Image-to-video, depending on the selected model | One required image; Seedance 2.5, 4 seconds, 720p preset, audio, 40 paid credits |
| Last frame | A visual destination; not the full motion between frames | Start/end-frame mode appears in its image-to-video form; confirm model support | No last-frame upload |
| Reference images | Appearance, subject or scene; upload order is not automatically a timeline | Omni Reference mode; count and compatibility depend on the model | The video form accepts only one first frame |
| Reference video | Movement to transfer to a reference subject | Kling motion control uses a video and a reference image | No arbitrary motion-reference video upload |
Sources: VidFlux image-to-video ↗ · VidFlux motion control ↗ · Current PromptToVisual form. Availability of a control does not guarantee exact identity, timing or endpoint matching.
View VidFlux image-to-video options. The H3 Max API examples below describe a separate external service.
Choose controls for the task
| Your task | Control | What to prepare |
|---|---|---|
| Keep a subject or look recognizable | Reference images | Identity or style material; not automatically timed anchors |
| Animate an existing still | First frame | One starting picture and a modest motion instruction |
| Connect two specific compositions | First and last frames | A plausible start and destination |
| Include a particular intermediate composition | First, middle and last frames | Three compatible pictures plus an interior time; inspect the H3 Max reference-to-video contract |
| Move the viewpoint around a scene | Camera trajectory keyframes | Camera poses along a normalized timeline, not three uploaded images |
First and last frames: choose start and destination
For H3 Max on fal, minimax/h3-max/image-to-video documents image_url for the first frame and end_image_url for the last frame. This is a useful contract to inspect when your task has two visual endpoints. A starting picture alone leaves the destination to the model and prompt; two endpoints still do not specify every intervening action.
First, middle and last: use the correct endpoint
The three-image contract is minimax/h3-max/reference-to-video. Use image_url, middle_image_url and end_image_url; middle_frame_time sets the intermediate time in seconds. A middle image requires both start and end images and native 480P or 768P output. Its time is rounded to the nearest frame at 24 fps and must fall strictly between the first and last frames of the requested duration. These constraints apply to that fal endpoint; they do not describe the PromptToVisual form.
Camera keyframes are a different input
minimax/h3-max/camera-controls takes an initial image_url and a camera_trajectory of poses. Each keyframe has time normalized from 0 to 1, azimuth and elevation in degrees, and distance in normalized scene units. Do not send middle_frame_time seconds as normalized camera time or treat three pose entries as three uploaded pictures.
Prepare inputs before paying for an attempt
Use images with a consistent subject, proportions and framing. Check that the intended movement can plausibly connect them within the chosen duration. Keep image preparation separate from motion instructions: label which picture is the start, the intermediate target and the destination. For reference-only inputs, describe the identity or style to preserve rather than assuming the upload order sets a timeline. These are preparation recommendations, not guarantees of exact preservation.
Plan a three-frame sequence
Imagine a five-second clip of a toy car crossing a tabletop. Plan a starting image at the left edge, an intermediate image near the center at 2.5 seconds, and a final image near the right edge. The intermediate time is inside the clip, and the pictures describe a plausible sequence. Use these positions to plan your inputs; they do not guarantee that the car follows an exact path. Choose a duration and resolution supported by your selected endpoint.
Common reasons a keyframe plan fails
A reference upload may influence appearance without fixing a moment. Incompatible poses may invite a visible morph instead of the intended motion. Too much action in a short interval can defeat a clear transition. A request can also fail validation when it uses the wrong endpoint, omits required companion images or mixes time units. Separate input validation from output quality: a successful request still needs visual inspection.
Where should I go next?
For three-frame control, read the external H3 Max reference-to-video contract linked below. PromptToVisual currently exposes only its Seedance single-first-frame workflow; it has no middle-frame, end-frame or camera-trajectory upload controls. Its four-second, 40-paid-credit option is suitable only when that narrower task fits. Ordinary editing-software keyframes animate properties on a timeline and are outside this guide’s main generation task. No Grok application capability is asserted here.
Sources and next steps
H3 Max reference-to-video contract →
H3 Max image-to-video contract →
H3 Max camera-controls contract →
Utopai X access guide · Keyframe control guide · FLUX 3 Video guide · Seedance limits · PTV single-first-frame alternative · All guides