The text to video workspace. The form sits on the left, the finished clip plays on the right.

The workspace keeps the form and the result side by side. You can adjust a setting and regenerate without losing the previous output.

Everything below happens in that left column, top to bottom. The order on screen is the order you should fill it in, because the model you pick decides which controls appear underneath.

Choose Text to Video. Changing the type rebuilds the form beneath it rather than navigating to a new page, so switching between modes keeps your place.

Text to Video is the only entry here that needs no input file at all.

Five generation types. Text to Video is the second entry.

The model list. FLUX 3 stops at 1080p and 20 seconds, MiniMax H3 at 2K and 15 seconds.

Those two numbers are limits rather than suggestions. Picking a model with a 1080p ceiling removes the 4K option from the resolution control further down.

Read the list before you commit. If the delivery needs 4K or a clip longer than 20 seconds, Seedance 2.5 is the only model here that can produce it.

Describe the movement, not only the scene. A prompt that names a subject and stops leaves the camera to the model, and the model will invent one.

Name the shot size, the direction of travel, and what stays still. A 30 second clip needs more of this than a 5 second one, because the model has to decide what happens in the middle.

The prompt field. The counter in the corner reads 0 / 2500.

Seven aspect ratios. The seventh tile, adaptive, hands the decision to the model.

Aspect ratio is a composition decision taken before generation. The model stages a subject differently in a vertical frame than in a widescreen one, so a vertical cut and a widescreen cut of the same idea are two generations.

The adaptive tile is worth using when you have no fixed delivery format and want the model to pick what suits the described shot.

Resolution and duration together drive both the wait and the credit cost, and they multiply. A 30 second clip at 4K is the most expensive combination available in the video workspace.

Test at 480p. Confirm the motion reads correctly, then raise the resolution once the prompt is settled.

Four resolutions. 720p is selected when the form loads.

A continuous slider, so any length between 4 and 30 seconds is available.

This is a slider rather than a set of fixed choices, which the older Seedance variants use. Seven seconds is available here and is not available on a model offering only 5, 10 and 15.

Cost scales directly with the number you set. Leave it at 4 seconds while you are still testing the prompt.

Mention the sound you want inside the prompt. "Rain on a metal roof", "a shutter clicks as the subject turns", and similar cues are read alongside the visual description.

Turning the toggle off lowers the credit cost. The exact figure updates above the Create button as soon as you change it.

One toggle. Audio is generated with the picture rather than added afterwards.

Up to 30 reference images. JPEG, PNG, WEBP, GIF and AVIF, to 300MB, minimum 300px by 300px.

Reference images anchor style, character appearance and scene composition. The AI Actor and AI Character sources pull a face you already built in the Studio, which keeps a person consistent across separate generations.

Previous Task is the fastest route when the anchor is something you generated an hour ago. It saves downloading a file and uploading it again.

A reference video supplies motion rather than appearance. Use it when the camera move or the action already exists in footage and you want the model to follow it.

Attaching reference videos changes how the generation is priced, because the length of the reference material counts toward the billed time. The Credits required line updates as soon as you add one, so read it before you press Create.

Up to 10 reference videos. MP4 only, to 1000MB each.

Up to 10 reference audio files. MP3 and WAV, to 20MB and 30 seconds each.

Reference audio guides the rhythm and character of the generated track. It is the section that makes music driven work practical, because the cut can follow a beat you supplied.

The three reference sections add up to the 50 asset capacity announced for the model. 30 images, 10 videos and 10 audio files.

Connect a Prompt node to a Text to Video node, then wire the output to a Video Viewer node. The Text to Video node shows a green "Prompt connected" line once the wire is in place.

An amber line reads "Connect a negative prompt node (optional)". A negative prompt is a second Prompt node listing what to keep out of the clip, and it is where exclusions belong.

The node reads ~4500 credits for 1080p at 5 seconds. The side panel carries the same four controls as the workspace form.

The finished clip opens full width in the viewer, with Download and Delete beneath it.

Click the output thumbnail on the node to open the clip full width. Download saves the file, Delete removes it from the graph.

The node keeps its result after it runs, so you can change one upstream setting and compare the new output against the previous one without rebuilding the graph.

The form is shorter than the text-to-video one. There is no aspect ratio control, because the source image sets the frame shape.

The three reference sections are absent too. In this mode the input image and the optional last frame are the reference material.

The image to video form. Credits required reads 4500 for 1080p at 5 seconds with audio on.

Image to Video is the first of the five generation types.

The form rebuilds immediately. An Image Source section replaces the aspect ratio tiles, and a Last Frame Image section appears below it.

Your prompt text survives the switch, so an idea written for text-to-video can be reused against a still without retyping.

Seedance 2.5 is the only entry in image-to-video that combines a 4K ceiling with a 30 second range.

Everything below this control is filtered by it. Choosing a different model changes which resolutions and durations remain available.

The image to video model list, with Seedance 2.5 at the top.

Five ways in. Upload, a previous generation, an AI Actor, an AI Character, or a URL.

This image sets the aspect ratio of the finished clip. Frame it in the shape you want the video to be, because there is no ratio control to correct it later.

A clean source helps. Clear subject separation, even lighting and few compression artefacts give the model more to track when it builds the motion.

This is the control to use for a continuation. Take the last frame of a finished clip, make it the first frame of the next one, and the two cut together cleanly.

Keep the two frames similar in subject position, lighting and shape. A large gap between them produces an abrupt transition in the middle of the clip.

Optional. Setting it fixes the composition the clip ends on.

The same prompt field, 2,500 characters, in both modes.

The model already has the appearance from your image. Describing the subject again wastes characters that should go on the movement.

Write what changes over the clip. The direction the camera travels, what the subject does, how the light shifts, and what holds still.

Resolution here is worth matching to the source. Animating a small image at 4K adds cost without adding detail the source never had.

Run the first pass at 480p and 4 seconds. That combination is cheap enough to test whether the still animates well at all.

The same four resolutions as text to video.

4 to 30 seconds, the same continuous slider.

A still that animates well for 5 seconds does not always hold for 30. Motion invented from a single frame drifts further the longer it runs.

Build up in steps. Confirm at 4 seconds, then 10, before committing credits to a full 30 second pass.

Generate Audio behaves the same way in this mode. The soundtrack is produced with the picture and follows what the prompt describes.

Press Create. The Credits required line above the button reads the exact cost, and the clip replaces the placeholder in the player when it finishes.

The audio toggle, on by default in both modes.

Image Upload and Prompt both feed the Image to Video node. This run reads ~3600 credits.

Wire an Image Upload node and a Prompt node into an Image to Video node, then send its output to a Result node. The generation node confirms both inputs with green "Image connected" and "Prompt connected" lines.

The Result node collects batch output, so a graph producing four variants gathers them in one place rather than four.

The output opens in the same full width viewer as the text-to-video graph, with Download, Delete and Close beneath it.

Reusing one Image Upload node across several generation nodes is the point of building this in a graph. One source frame, several models or several settings, side by side.

The finished clip, with Download, Delete and Close below the player.