How to write a Seedance 2.5 prompt
A prompt for a video model like Seedance 2.5 is 13 labelled sections written one after another. HFLUX PROMPT 101 fills the 14 fields behind those sections from a single line of brief, and you can edit the fields yourself.
The 13 sections
Use the labels as they are, in this order, and write the content after each one.
| Section | What goes in it |
|---|---|
| GLOBAL STYLE | Genre and look, shutter and grain, the treatment everything else has to match, the colour split |
| SCENE | One line: what happens, where, in what mood |
| CHARACTERS | Face, hair, build and wardrobe of everyone on screen, and a statement that nobody else appears |
| LOCATION | The space and its props, kept separate from the people |
| FIRST FRAME AND SPATIAL BLOCKING | Exact positions and facing the moment the clip begins |
| FORMAT MODE | How many shots, where the hard cuts land, the total runtime |
| OPTICS | Field of view, aperture, depth of field, flares |
| CAMERA | Lens height, rig (handheld, dolly, crane, locked-off), how it behaves across the shots |
| ACTION TIMING | The shot-by-shot breakdown — one shot per line |
| PHYSICS | How fabric, hair, smoke, water and mass behave |
| LIGHTING | The motivated sources, their direction, how light falls |
| AUDIO | Ambience and specific effects, which spoken lines exist and what must not be there |
| POSITIVE LOCKS | What stays identical in every frame and what must never appear |
The official guide notes that skipping a section tends to fail in a specific, predictable way If a field is hard to fill, a short line beats leaving it empty
One shot per line, Hard cut between them
ACTION TIMING is written differently from the other sections. One shot per line, ending with Hard cut — except the last one.
Timings run continuously with no gaps and add up to the total runtime. Instead of a vague "and then", say what the cut lands on.
Text-to-video and image-to-video
T2V means you write a prompt and get a video. I2V means you upload a photo and the model animates it. Pick I2V when a specific person, product or object has to come out the same; pick T2V when there is no reference.
The notation differs. With a reference photo you declare the reference, and then sound effects are not marked with angle brackets — the same symbol must not carry two meanings. The declaration itself is model-specific: 2.5 uses <Name> corresponds to @Image 1. while 2.0 uses Define ... in Image 1 as <Subject_1>.
Pick the runtime first
| Model | Runtime |
|---|---|
| Seedance 2.5 · Wan 3.0 | 5 – 30 s (1 second steps) |
| Seedance 2.0 · 2.0 Fast · 2.0 Mini · MiniMax H3 | 5 – 15 s |
| Kling 3.0 Turbo · v3 · v3 Omni | 3 – 15 s |
| Gemini Omni 1.1 | 10 s |
The shot breakdown is divided to fit it. Values such as aspect ratio and duration are not written as sentences in the prompt — they are set on the page or through the API, and the requested duration only guides how the shots are divided.
Why the prompt is in English
Everything that reaches the video model is written in English. The official Gemini Omni 1.1 Flash documentation, for instance, states that English only is fully supported.
Two things keep their original script — words that must appear on screen (a sign, a caption, a title) and lines a person actually speaks. Both go in straight double quotes, and a spoken line names its language first.
Every field also carries a one-line description in your interface language, so you can read what the English text says. HFLUX runs in 9 interface languages: Korean, English, Japanese, German, Portuguese, Spanish, French and Chinese (Simplified and Traditional).
Using PROMPT 101
- Pick an example or write one line — there are 12 examples
- Choose the target video model, the branch (T2V or I2V), the runtime and the aspect ratio
- Press Generate prompt and the 14 fields fill in — a 35 credit base fee, plus an amount for the length of your brief
- Edit the fields — any field can be edited by hand
- Press Finish prompt and the 13 sections join into one block
- Send it to the Video tab, or copy it, download TXT, or save JSON
You can also start from blank fields A saved JSON can be loaded again to keep editing
It is a node in the canvas too
The HFLUX node canvas has a PROMPT 101 node. Set the brief, the target model and the branch on the node, generate, and the finished prompt flows straight into the next node. The canvas has 37 node types in all — prompt, video, image, storyboard, background removal, object removal, depth, pose, 3D studio, speech, sound effects, music and more — with 68 presets arranged in seven groups.
Changing only the camera, light and colour of an existing clip
To rework a clip you already have, you do not write a new prompt — you use RECAST VISION. It changes the camera (shot size, angle, move, lens), the lighting (direction, colour, intensity, diffusion, up to two sources) and the colour (palette) of the uploaded clip, applied to the whole clip at once. Billing follows the length of the source.
AI VFX — stacking the control maps
The canvas presets include an AI VFX group of eight. Six of them carry the names of the real VFX stages: roto (subject and background apart), cleanup (erase, fill, finish), redo from depth and pose, take the pose only, previs to final shot, and all three control maps. They stack background removal, object removal, depth map, pose, 3D pose and 3D STUDIO into a single shot. The other two (Change the world and Compositing pipeline) are examples that open canvases HFLUX actually built, results included.
Control maps occupy the reference-video slots. Seedance 2.0 allows up to 3 of them, 15 seconds combined, object removal and keep-only-object run on up to 10 seconds of video, and background removal, depth map and pose run on the first 30 seconds.
Frequently asked questions
How do you write a Seedance 2.5 prompt?
Write 13 sections in order, each with its label: GLOBAL STYLE, SCENE, CHARACTERS, LOCATION, FIRST FRAME AND SPATIAL BLOCKING, FORMAT MODE, OPTICS, CAMERA, ACTION TIMING, PHYSICS, LIGHTING, AUDIO and POSITIVE LOCKS. In HFLUX PROMPT 101 a single line of brief fills the 14 fields behind those sections, and you can edit any field before joining them into one prompt.
How are shots written in a Seedance prompt?
One shot per line in the ACTION TIMING section. The shape is Shot 1 (0-3.5s): followed by the shot size and angle and what happens, ending the line with Hard cut. The last shot has no Hard cut after it. Timings run continuously with no gaps and add up to the total runtime.
What does PROMPT 101 cost?
Having the AI fill the 14 fields costs a 35 credit base fee, plus an amount for the length of the brief you wrote. You can also load one of the 12 examples or start from blank fields and edit any field, then copy the finished prompt or download it as TXT.
Why is the prompt in English?
Everything that reaches the video model is written in English. The official Gemini Omni 1.1 Flash documentation, for example, states that English only is fully supported. Two things keep their original script: words that must appear on screen, and lines a person actually speaks. Both stay in straight double quotes. Each field also carries a one-line description in your interface language.
What is the difference between text-to-video and image-to-video here?
Text-to-video means you write a prompt and get a video; image-to-video means you upload a photo and the model animates it. The notation differs: with a reference photo you declare the reference, and in that case sound effects are not marked with angle brackets, so the same symbol never carries two meanings. PROMPT 101 shows the notation guide for whichever branch you picked.
Which video models can it target?
Eleven: Seedance 2.5, Seedance 2.0 with 2.0 Fast and 2.0 Mini, two Wan 3.0 models, Kling 3.0 Turbo, v3 and v3 Omni, MiniMax H3 and Gemini Omni 1.1. The official grammar for the model you pick is included, so the prompt comes out shaped for that model.
What runtimes can I choose?
Whatever the model supports. Seedance 2.5 and Wan 3.0 run from 5 to 30 seconds in one second steps, the Seedance 2.0 family and MiniMax H3 from 5 to 15, Kling 3.0 from 3 to 15, and Gemini Omni is 10 seconds. The shot breakdown is divided to fit the runtime you chose.
Can I generate the video straight from the prompt?
From the finished screen you can send it to the Video tab and the prompt goes in as it is. You can also copy it, download it as TXT, or save it as JSON and load it again later to keep editing. The node canvas has a PROMPT 101 node as well, so it can feed other nodes directly.