[Successfully animated 16 people]: A summary of Seedance 2.5 prompt notation rules, differences from 2.0, and basic templates.
*I am revising and adding to this article as I verify information, so 'article updates' will continue for a few days.
Latest/Updated: Saturday, August 8, 2026, 9:00 AM
Article Published: Friday, August 7, 2026
Seedance 2.5 is now available!
Seedance 2.5 has just become available on Runway, and surprisingly, it is unlimited.
*Only one item can be in the queue (2.0 allowed two).
*Generations are limited to 15 seconds.
I honestly thought unlimited 2.5 would be impossible, but it exceeded my expectations in a good way.
...So, I have decided to change my strategy immediately...
I am switching the video generation model for my AI drama production to Seedance 2.5.
I had given up on using 2.5 because it is expensive (I intended to stick with 2.0), but if it is unlimited, I have to switch!
In any case, this is a race against time until the end of this month (when unlimited access will be completely discontinued). 2.5 costs 450 credits for 720p/15 seconds, which is only 21 generations even with the Max plan (which provides 9500 credits).
Since the specifications have changed significantly, it looks like I will need to improve my workflow, but first I want to check the official information that has been released.
Seedance 2.5 exceeded 2.0 more than I imagined!
I am still in the early stages of verification and will proceed little by little, but I will document the visual expressive power and other features that clearly surpass 2.0.
Laws of Physics
The area where the 'difference is clearly visible' is the laws of physics.
The following is a scene from an MV where a large electric fan is running in front of the person. In other words, it is a scene where they are performing while being hit by intense wind.
In 2.0 (top), there are two mutually inconsistent states of motion existing simultaneously: 'thin strands of hair being hit by very strong wind' and 'large locks of hair barely being affected by the wind.' Simply put, it is 'quite unnatural'.
In 2.5 (bottom), the entire volume of hair, including not just the tips but also from the head to the cheeks and around the neck, is spreading outward. While local thin strands are moving, the large locks of hair beneath them are also being subjected to the same external force.
2.0:
It looks like individual movements are occurring for each hair bundle2.5:
It looks like the entire hair is reacting to the wind as a single deformable object
Playback time: 34 seconds
Seedance 2.5
In 2.0, it is highly likely that 'maintaining the reference image' and 'moving the hair' were in conflict. The MV I created in 2.0 will have to be remade in 2.5...
Reference fixation, emotional expression, and continuous movement
2.5 is clearly superior for continuous movement without breakdown.
In the case of the video below, there is a continuous sequence of:
[A: walking] -> [A: crouching] -> [A: picking up a pebble] -> [A: throwing it into a puddle] -> [B: mud splashes on face] -> [B: getting angry] -> [A: apologizing] -> [C/D: laughing]
as a continuous movement.
Playback time: 15 seconds
Seedance 2.5
2.5 succeeds no matter how many times I generate it, but 2.0 fails half the time after 6 generations. Failure means the same person appearing, roles changing, or the order of alignment changing, etc.
Animating 16 main characters
I tried animating using references for 16 main characters appearing in an AI drama. In 2.0, the limit was 9, so it was 'up to 9 people', but in 2.5, it has been expanded to 50.
First, I was surprised by the high success rate.
*It decreases if you don't add reinforcement prompts for the references.
Playback time: 43 seconds
Seedance 2.5
Since the 2.5 prompt becomes a massive amount, I will only post the part for the 16 characters.
Reinforcement prompts for 16 people (part of the whole):
Characters:
16 people: Hikari (@Hikari_A), Akari (@Akari_A), Mako (@Mako_A), Risa (@Risa_A), Johnny (@Johnny_A), Kurosaki (@Kurozaki_A), Yamada (@Yamada_A), Ariel (@Ariel_B), Marcus (@Marcus_A), Evan (@Evan_A), Becky (@Becky_B2), Hitomi (@Hitomi_B2), Claire (@Staff_Claire), Martin (@Staff_Martin), Nia (@Staff_Nia), and Maya (@Maya_A1).
All 16 people speak Japanese.
Appearances of the 16 people:
Hikari (@Hikari_A) has black hair in twin buns with bangs. She wears a shiny silver down jacket, a bright pink top, a short pink skirt, blue and pink leggings, long-pile pink leg warmers, and pink shoes.
Akari (@Akari_A) has black hair in high twin tails with colorful hair clips. She wears a pink sleeveless slim top, flashy graphic-print wide pants, white fur leg warmers, and pink platform shoes.
Mako (@Mako_A) wears glasses and has a black ponytail with colorful hairpins. She has clear skin with no moles. Her outfit consists of a lime green mesh long-sleeve top over a pink velvet camisole. She wears a denim mini skirt with "COLORS" spray-painted on the front in a pink and light blue gradient. She also wears a pink fur waist pouch, white fur leg warmers, and pink platform sneakers.
Risa (@Risa_A) has a vibrant blue short bob with bangs and doll-like makeup. Her outfit is a white mini dress with a small floral pattern (front buttons, lace details). She wears a white leather heart choker, a silver glittery pouch, white fingerless gloves, and white lace-up platform long boots.
Johnny (@Johnny_A) has straight, shoulder-length black hair, parted to one side, and is clean-shaven. He wears a black leather long coat with a high collar, silver studs on the lapels and cuffs, floral embossing on the shoulders and chest, a black vest underneath, black pants, and black leather boots. He wears layered silver necklaces, a cross pendant, dangling earrings in both ears, and brooch and chain decorations on his coat.
Kurosaki (@Kurozaki_A) is an elderly man with white hair, wearing black work clothes.
Ariel (@Ariel_B) has a pastel pink short bob with bangs covering one eye and bright gray eyes. She wears a dark navy FBI field jacket with large yellow "FBI" lettering on the chest and sleeves, a black tactical vest underneath, dark gray slim pants, and black ankle boots.
Marcus (@Marcus_A) is a broad-shouldered, muscular Black man. He has dark brown skin, a shaved head, deep forehead wrinkles, thick black eyebrows, deep-set dark brown eyes, and a short black full beard connected to a mustache. He wears a navy windbreaker with yellow "FBI" lettering, a black tactical vest with large white "FBI" lettering, a plain black crew-neck shirt, dark navy jeans, and brown leather lace-up work boots.
Evan (@Evan_A) is a tall, muscular white man. He has an angular face, a square jaw, bluish-gray eyes, straight brown eyebrows, and short light brown hair styled upwards. He wears a navy windbreaker with yellow "FBI" lettering, a black tactical vest with large white "FBI" lettering, a faded blue chambray shirt, dark navy jeans, and brown leather lace-up boots.
Becky (@Becky_B2) has a round face. She has a short, straight bob cut at the jawline. Her hair color is bright aqua blue. The bottom half and tips of her hair are dark turquoise blue. She has short, straight-cut bangs above her eyebrows. Her hair is straight. She wears a dark navy FBI field jacket with large yellow "FBI" lettering on the chest and sleeves, a black tactical vest underneath, dark gray slim pants, and black ankle boots.
Hitomi (@Hitomi_B2) has a shoulder-length black straight bob. It is center-parted with no bangs. The hair length is from below the jaw to near the collarbone. The ends are slightly tucked inward. She wears a dark navy FBI field jacket with large yellow "FBI" lettering on the chest and sleeves, a black tactical vest underneath, dark gray slim pants, and black ankle boots.
Claire (@Staff_Claire) wears a dark navy tailored jacket, a white button-down shirt, and dark navy slim pants. She wears black pointed-toe pumps.
Martin (@Staff_Martin) wears a well-worn brown leather jacket, a plain olive-gray T-shirt, and faded dark denim pants. He wears brown work boots. The jacket has scuffs, color unevenness, and wrinkles, giving it a texture of having been worn for many years.
Nia (@Staff_Nia) is a young Black woman with shoulder-length, tight black curly hair, center-parted. She has warm brown skin, natural makeup, and a smile showing white teeth. She wears a patchwork denim jacket with plain square patches in blue, red, yellow, and green, and silver buttons. Underneath, she wears a yellow T-shirt with thin blue horizontal stripes. She wears light blue high-waisted straight denim jeans and white high-top canvas sneakers.
Maya (@Maya_A1) has black hair center-parted and smoothed flat against the scalp at the crown, with thick rope-like twists wrapped into large, fist-sized round loops positioned on the left and right sides of her head from the temples to ear height. She wears flower-shaped stud earrings made of small gold beads, a form-fitting white short-sleeved crew-neck cotton T-shirt, high-waisted wide-leg beige cargo pants with side pockets with flaps, and pointed snakeskin-patterned heels.
This is just the amount for the characters alone.
The entire prompt is three times this length. I think you can understand why I say, 'It is impossible for a human to write this prompt (it is hellish work).'
Seedance 2.5 Prompt Notation Rules
As always, we will refer to the official user guide.
You must always start with the user guide for a new model, but in the end, the personal rules you build through trial and error afterward are the strongest.
However, you cannot make improvements without starting from the basic specifications, so we start with the official one.
Official Guide:
Level 0
This is the same as 2.0, so it remains as it was.
1. Do not write generation parameters in the prompt.
Configurable parameters such as aspect ratio and duration should be specified on the generation page or via the API. There is no need to write them in the prompt body (they become noise information).
2. Write the state that can be directly seen/heard, not abstract words.
Words like 'tense' or 'warm atmosphere' only convey a direction and leave too much room for interpretation in acting and visual composition.
For example, for emotional direction, describe observable changes such as eye movement, tension in the eyebrows, breathing, hand movements, or the foreground blurring while the background becomes sharp.
3. The model will fill in any parts left ambiguous.
If you do not write the role of reference materials, which image corresponds to whom, or the end state of a scene, the model will guess on its own, leading to results that deviate from your intent.
Conversely, the more precise specifications you provide, the more stable the results will be.
Scripting the Prompt
Because Seedance 2.5 requires prompts to be written thoroughly, the limit for Runway has been expanded to 15,000 characters.
Runway previously only allowed 3,500 characters, so it was difficult to keep things short. Even with Seedance 2.0, there were times when it exceeded 4,000 characters.
In AI drama production, you will extract from the script to create a base prompt, so with 2.5 allowing up to 50 references, the amount of text becomes quite large.
If a human writes the prompt, it becomes a hellish task (like proofreading an entire book), which is time-consuming and costly, so I recommend leaving it to Codex (or Claude Code).
Level 1 is generally what we have been doing so far, so it is not expected to require major modifications. Although I don't know for sure as I haven't verified it yet.
Be careful with audio notation. If you use the symbols incorrectly, it will likely be a waste of effort.
As I noted before, once you reach the Seedance 2.x class, it is 'almost impossible for humans to write prompts,' so you will need to provide Codex with information on the following specifications and have it set up a prompt generator using your secret sauce (your past data).
Level 1
Define who, where, what they are doing, and what the final outcome is in a way that the model can observe.
The basic structure is as follows.
Subject
+ Action or event
+ Scene and environment
+ Visual style
+ Camera work/cut
+ Audio
1. Determine 'who does what'
Bad example:
Cinematic and moving footage. Beautiful lighting. Tense atmosphere.
In this text, it is not decided what specifically will happen.
Good example:
A young potter lifts a blue cup from the potter's wheel in a workshop at dawn and places it in the center of a wooden shelf.
First, decide on the following four items.
Who or what is the protagonist
Where they are
What they do
What is visible at the end
2. Place only one major change in each stage
For videos containing multiple events, divide them into stages.
Place the following three items in each stage.
Initial state
Key events
End state
The "End state" is particularly important.
For example, just saying "a woman wraps a bouquet" leaves room for her to drop the bouquet halfway or for another person to hold it.
Write it down to the directly visible state as follows.
End state: The wrapped bouquet is placed flat in the center of the workbench, with the green ribbon bow facing the camera.
For long-form generation in Seedance 2.5, it is recommended to set only one major state change for each stage and explicitly state the observable condition at the end of the stage.
3. For reference materials, write not only "what to use" but also "what not to use"
Just uploading reference images is not enough.
Explicitly state the attributes that each material provides to the model.
@Image 1 defines only the <woman>'s facial features, hairstyle, and outfit.
Do not use the image's background or composition.
@Image 2 defines only the <cafe>'s interior layout, furniture, and light from the window.
Do not use the people in the image.
@Video 1 defines only the action and tempo of the <woman> lifting a cup.
Do not use the appearance, outfit, or background of the person in the video.
Since there is a possibility that the people, background, or composition of the reference materials may be unintentionally mixed in, specify both the attributes to use and the attributes to exclude.
Do not have the model read text labels in images or guess the roles of the materials.
4. Individually name and map people, props, and backgrounds
If multiple people appear, avoid ambiguous designations like the following.
@Image 1 through @Image 4 define 4 characters.
This does not confirm which image corresponds to which person.
Correctly, map them one-to-one.
<Clerk> corresponds to @Image 1.
Use only appearance, hairstyle, and uniform.
<Customer> corresponds to @Image 2.
Use only appearance, hairstyle, and outfit.
<Red bag> corresponds to @Image 3 and is owned only by <Customer>.
<Cafe> refers to @Image 4.
Use only interior layout, furniture, and lighting; do not use the people in the image.
By separating not only the appearance of the people but also their outfits, positions, actions, lines, and ownership of props, it becomes easier to suppress character swapping.
5. More reference materials are not necessarily better
The specification of "up to 50" does not mean that using 50 is recommended. As the number of materials increases, stability decreases.
When showing the same person or product from multiple angles, rather than combining them into a single collage, prepare the front, side, and back views as independent images. Then, explicitly state that "all are the same person" or "all are the same unit."
@Image 1 defines the front of the same lamp.
@Image 2 defines the left side of the same lamp.
@Image 3 defines the back of the same lamp.
All three images define one single lamp.
Throughout the entire video, there is exactly one lamp.
*I can't say for sure without verifying this...
In a scene with 7 main characters, the references would total 21.
Update (August 8):
I tried animating 16 characters. As before, a single character sheet seems to work fine.
Adding 16 references certainly makes the screen look incredible. I will try verifying up to 25 or 30 people next.
6. Fix consistency with specific items
"Maintain consistency" is too vague.
I will list what needs to be maintained.
[Maintaining Consistency]
- Do not swap the identities of the two characters.
- Keep the number of people at two throughout the entire video.
- Do not change costumes.
- Only the <customer> owns the red bag.
- Do not change the left-right orientation of the table.
- From the camera's perspective, the <clerk> is on the left of the frame, and the <customer> is on the right.
- Do not duplicate or split people or props.
7. Do not think of timestamps as frame-by-frame commands
Timestamps are for allocating time to events.
0-5 seconds: The woman approaches the table and stops in front of the cup.
End state: The woman's right hand is next to the cup handle.
5-10 seconds: The woman picks up the cup and takes a sip.
End state: The cup is held in the woman's right hand.
The important principles are as follows:
Make time ranges continuous
Do not overlap time ranges
Do not pack too many actions into a single second
In normal storytelling, prioritize stage structure over second-based specifications
Timestamps are not strict edit points
Specifying things like "complete 3 actions in 1 second" may lead to excessive cuts or skipped movements.
8. Use different notations for audio
Prompts can be written in natural language alone, but if you want to clearly separate audio elements, use the following notation.
Music
Notation: ( )
Example: (Quiet piano music plays in the background)
Sound Effects
Notation: < >
Example: <A bell rings in the distance>
Dialogue
Notation: { }
Example: {Good morning}
Subtitles
Notation: 【 】
Example: 【Chapter 1: Departure】
For dialogue in languages other than Chinese, specify the language before the dialogue.
A woman says calmly in Japanese: {Sorry to keep you waiting}
To specify a language or regional variant more strongly, use the following order:
Language of dialogue
+ Regional variant or accent
+ Manner of speaking
+ Speaker
+ {Dialogue}
9. Define start and end frames separately
When using start and end images, write the role of each image individually.
@Image 1 is the start frame.
Define the composition, character position, posture, prop state, and camera direction at the start.
@Image 2 is the end frame.
Define the composition, character position, posture, prop state, and camera direction at the end.
It is recommended not to combine them into a single sentence as shown below.
@Image 1 and @Image 2 are the start and end frames.
Start and end images must have the same aspect ratio.
If the ratios differ, the end image may be stretched. Additional reference images can supplement character or prop attributes, but they must not change the composition defined by the start and end images.
10. In video editing, 'things to maintain' are more important than 'things to change'
When editing an existing video, first define the source video as the sole editing master.
[Editing Goal]
Edit @Video 1.
Change the blue light on the right wall to a warm orange only in the 4-7 second range.
[Role of Source Video]
@Video 1 is the sole editing master.
Defines the characters, actions, composition, camera movement, occlusion relationships, audio, and sequence of events.
[Editing Scope]
Change only the color of the light on the right wall and the area illuminated by that light.
[Things to Maintain]
Maintain character identity, costumes, expressions, positions, actions, interior structure, camera movement,
dialogue, and ambient sounds.
The following four points are necessary when editing:
Sole editing master
Target to change
Time or area to change
Content that will not be changed
In video editing, the aspect ratio and approximate duration of the input video are automatically maintained and cannot be changed individually. There may be a difference of up to approximately 0.3 seconds in the output duration.
Basic Template
[Generation Goal]
<What kind of video to create>.
The central subject is <Subject Name>.
The main event is <Short summary of the event>.
[Reference Material]
@Image 1 defines only the <face, hairstyle, outfit> of <Subject Name>.
Do not use the background or composition of the image.
@Image 2 defines only the <spatial layout, furniture, lighting> of <Location Name>.
Do not use the people in the image.
@Video 1 defines only the <action or camera movement>.
Do not use the appearance, outfit, or background of the people in the video.
[Initial State]
<Initial state of people, props, background, and camera>.
[Main Event]
<One continuous main action performed by the subject>.
[End State]
<Position of people, position and owner of props, final visible state on screen>.
[Visuals]
<Shot size, camera position, camera movement, lighting, screen texture>.
[Audio]
<Ambient sound>
<Action sound effects>
<Speaker says in Japanese: {Dialogue}>
(Music)
[Maintaining Consistency]
Maintain <person identity, number of people, outfits, number and ownership of props,
left-right orientation, background layout, camera axis>.
Do not duplicate or split people or props.
Headings like [Generation Goal] are not special syntax like phonetic symbols, but a structure for organizing prompt information.
Diagnosis when failure occurs
People get swapped
Possible causes:
Each image and person are not mapped one-to-one
Outfits, actions, positions, and dialogue are not separated
Using references unnecessary for the scene
Not specifying the owner of props
The same object increases in number
Possible causes:
Not specifying the quantity
Not clearly stating that images from multiple angles are the same object
Not writing 'exactly one throughout the entire video'
Actions are skipped
Possible causes:
Packing multiple state changes into a single stage
Too much movement for the given duration
End state is not defined
The end image is distorted
Possible causes:
The aspect ratios of the start and end images differ
Additional reference images are overriding the composition of the end frame
The entire video changes during editing
Possible causes:
The source video is not defined as the 'sole editing master'
The area or duration being edited is too broad
Content to be maintained is not listed
The number of objects is not specified
Emotional expression is weak or excessive
Possible causes:
Only abstract emotion words are used
Changes in eyes, eyebrows, mouth, breathing, gaze, and hands are not specified
Too many details are packed into a single emotion
From level 2 onwards, the difficulty is quite high, so I will verify things as I go. After all, it only became available '2 hours ago,' so nothing has been verified yet.
Update:
Resolved
There was an option to switch to MP4.
I was told about it!
This is a Runway-specific issue, but for some reason, only Seedance 2.5 downloads files in 'mov' format.
Moreover, it seems to be a special mov format, causing errors in apps other than QuickTime Player and tools like Topaz Astra....
I was able to upload the downloaded video file to ChatGPT and convert it to a standard MP4, but since that's tedious, I automated it with Codex. I set it up so that it converts automatically just by placing it in a specific folder.
The work in Codex required about 20 approval operations and took about 3 hours, but I will use this system for conversion until Runway's specifications change.
Related Magazine:
Updated: Friday, August 7, 2026 / Published: Friday, August 7, 2026