Start with the task. Then choose the model

Your image is ready. You know how the scene should develop after that frame. But you face another list of models: which one accepts this material, where is the right mode, and why does one name have several versions? We changed the order of that decision and moved the essentials to the top of the Video node.

“What do you want to do?” is a better first question

The selector has eight roles, from creating with a description to working on an existing video. Recast characters replaces performers in the original scene; Avatar transfers motion to the character in an image. The tasks below follow the menu order; choose one to jump to its walkthrough.

The interface screenshots were captured in the local studio using a sample project. These are working menus and a real node, not a drawn mockup.

Compatible models first. Explanations for the rest

After you choose a role, models are ordered by support for that operation and the type of material connected. Matches come first. Other models available to your account stay visible, with a reason why they do not fit this choice, such as an unsupported operation or an input that this role does not use.

Fewer lines on the node. More useful information at a glance

The role sits beside your node’s name, with the model and current settings below it. Explicitly choosing a role also adjusts its inputs: Create from a description keeps scene and prompt. Unused empty inputs disappear; materials already connected stay connected when you change the role.

Role icon

Click to choose a different action.

Model name

Opens the models for the current role.

Timer and other icons

Open a specific setting: duration, resolution, variant, or audio, where supported by the model.

The selected duration is not the model’s limit

The node screenshot has eight seconds selected. For video from a starting frame, Grok Video 1.5 supports up to 15 seconds, while Seedance 2.5 supports up to 30. Click Compare and select up to three models to see the difference. The view shows available information about maximum duration, resolution, and inputs for the chosen role and variant.

A variant is not a universal quality score. Seedance 2.0, for example, offers Draft, Fast, and Full; FLUX 3 generation offers Draft and Full. FLUX 3 Draft has a fixed 720p output. Changing the variant can change the available settings; matching variant names do not make different models equivalent.

Seconds mean different things too: for generation, the length of the new clip; for editing, possibly “same as source video”; for extension, the length of the added segment, not the whole film. Resolution sets frame size, while a variant selects a version within a model. Fixed parameters do not get a separate selection button in the compact header.

The estimated cost of the selected run now appears in the selector and updates when settings change. The RU badge is the owner’s listening assessment of Russian speech, not a pronunciation guarantee for every clip.

On the node: what I chose. In comparison: what I can choose.

Every role, explained through a task

Each walkthrough tells you what to connect, which role to choose, and what result to request. The images and clips come from our earlier projects, not a new model benchmark. The prompts are separate examples of intent, not promises to reproduce the shown results.

02Your image defines the first frame of the video

What to connect

A starting image + what happens next

What you get

A clip that starts from your frame

Choose “Video from a starting frame”: this is image-to-video. Connect the image to the first-frame input, then choose a model and duration. The image sets the initial scene; the prompt describes what happens next: a character’s actions, a developing scene, a transformation, or camera movement. Simply animating a picture is one possible use, not the definition of the role.

Example prompt

The camera slowly pushes in. The cucumber tilts his head slightly toward the tomato; she shifts her gaze toward him. Preserve the composition and soft lighting. No dialogue.

A first frame does not always rule out additional references. Kling, for example, accepts character Elements in this same role: the opening image sets the scene’s start, while Elements help describe the characters. This is a capability of that model within the role, not a rule for every image-to-video model.

Not one frame, but an opening and an ending

In the same Video from a starting frame role, choose a model that supports a last frame, such as Seedance 2.5. The first image sets the opening; the second sets where the scene should arrive. Connect them to the “first frame” and “final frame” inputs, then describe the transition. These are endpoints, not simply two character references.

Example: first and last frame →

Need to control the middle of the clip too? →

03You have a character and an object and need a new scene

What to connect

Character, object, motion, or audio references + a prompt

What you get

A new scene guided by the chosen references

Choose “Create from references”: this is reference-to-video. These images do not have to become the opening frame: they guide who or what should appear in a new scene. Connect the materials you need and give each a clear job in the prompt.

Example prompt

Use the selected character’s reference for their appearance, and the Key image for the artifact’s shape. The character notices the Key on a table; the camera racks focus from the face to the object. Do not add the other characters to the shot.

Not every model accepts everything together. Images, video, and audio can only be combined in an operation that supports them. An audio input does not by itself mean that audio alone is enough to start a run.

Image + video + audio, in one run

Seedance 2.5 accepts all three together under Create from references: an image guides the character or object, a clip guides motion or atmosphere, and audio provides a sound reference. This is one combined operation, not three generations in sequence. Input video and audio share a 30-second total budget; that is separate from output duration. Explain each file’s purpose in the prompt; the new shot does not have to reproduce a frame from the reference clip.

Example prompt

Use the image for the Key’s shape, the video for the slow camera push-in and drifting smoke, and the audio for rhythm and atmosphere. Create a new shot of the Key floating above a wrecked starship. Do not copy characters from the reference clip. No speech.

Only have audio: can you still create a video?

Yes: choose Seedance 2.5 and Create from references, connect the audio, and describe the visuals. This operation does not require an image or a video. Seedance 2.0 is different: you must add an image or video alongside the audio. Here, audio-to-video is a use of the references role, not a separate button available for every model.

Example prompt

Create visuals for the audio reference: glowing waves travel across a dark lake, their movement following the music’s rhythm. One continuous shot, without characters or speech.

This is not the output-audio switch. An audio reference is an input; “model generates”, “source audio”, and “no audio” control the output soundtrack. An audio reference by itself does not promise precise character lip-sync.

04Your plan is already broken into frames: make one clip

What to connect

Several frames in order — one storyboard sheet, or key frames each pinned to its own second

What you get

A single clip that follows those frames

Choose Shoot from frames when you know several successive moments of the video: she opens the door, pauses on the threshold, then steps inside. You request one clip following that sequence. How to prepare the images depends on the model; the selector explains the difference before you run.

Need to control the middle of the clip too?

Choose FLUX 3 under Shoot from frames. Connect the opening image to first frame, intermediate images to “key frame” inputs, and an ending image to “final frame” if needed. Set times for intermediate points: a helicopter at the opening, the transformation at three seconds, and a robot at the end. Up to 10 frames are supported, including the first and last. These are timed visual targets, not a collection of character appearances.

How to place video keyframes in time →

You have a finished storyboard sheet

For Seedance 2.0, Seedance 2.5, Grok, and other models with reference inputs, prepare one image containing the panels in order and connect it to an element input. You can use the Storyboard Sheet generator from the library. Ask the prompt to follow the panels in sequence. Panels have no individual timestamps; connecting several images does not automatically assemble them into a sheet.

Kling needs a starting frame alongside the sheet

Connect the storyboard sheet to an element input and supply the required first frame separately. The sheet takes one Element slot. The starting image defines the opening, and the sheet guides the following moments; individual panels cannot be pinned to a second here.

“They do it differently — read the caveat” explains how frames are supplied, not how beautiful the result will be. Timed keyframes and a storyboard sheet provide different control over a scene.

Example prompt

Follow the storyboard frames in order: she opens the door, pauses on the threshold, steps inside. One continuous shot, no cuts, no on-screen text.

Try it on your canvas

1. Add Video using “+”, or drag a wire from source material into empty canvas space and choose Video.
2. Choose a role, then a model. Open comparison here without leaving your canvas.
3. Review duration, variant, resolution, and audio. Available controls depend on the chosen operation.
4. Create the node or apply the selection to an existing one. Review its inputs and cost estimate before running.

When creating from a wire, you still choose the role. The selector then uses the material type to order models; it does not automatically guess the task.

The update is available to all studio users. The models have not become identical, and we have not started choosing for you. Their differences are now presented in terms of what you want to do.