Seedance 2.5 Long Video · One Prompt, 3 Minutes of Film?

Quick Summary

• Credit: The embedded video, the Tide Keeps short, the original prompt, and the test belong to JSFILMZ. This article uses the provided transcript and official Dreamina pages to separate demonstration, specification and analysis.
• The Seedance 2.5 Long Video beta officially supports generations between 5 and 180 seconds on Dreamina. The standard mode continues to be described with a maximum of 30 seconds.
• The test is not just text to video. JSFILMZ loads four character sheets, links each name to the corresponding reference, and sends a 3,721-character multimodal prompt in Chinese.
• The English version had approximately 7,000 characters. The Chinese translation helped stay under the 5,000 character limit shown in the interface, but the creator himself acknowledges that the translation is not semantically identical.
• The generation was set to 16:9, 180 seconds and 480p. The screen showed 20,000 credits in 480p and 40,000 in 720p. The rendering took approximately 25 to 30 minutes.
• The result covers three scenes, preserves the characters quite well and delivers dialogue, title, inserts and context images. For a first narrative draft, the density of information is impressive.
• The flaws are structural: inverted diner orientation, photograph shown to the camera instead of the other character, omitted actions, approximate interaction with props, slow pace and an important church almost out of the frame.
• The best reading is an assembly prototype and not a final film. A single request can yield three usable minutes of review, but spatial continuity, blocking, props, timing, and resolution still require iteration and post-production.

Credit and state of the evidence

The embedded test was published by JSFILMZ. The short, the narrative universe, the character sheets, the original prompt, the render and the video observations belong to the creator. The credit for the experiment and the decision to show the defects, instead of just the final result, belongs to him.

The analysis crosses the full transcript with Dreamina's official Seedance 2.5, Long Video Generator and How to Use Seedance 2.5 pages. The search was updated on August 4, 2026.

Durations from 5 to 180 seconds, standard mode up to 30 seconds and the announced limit of 50 references are official specifications. Quality, prompt, credits, render time and defects described in this test are observations by JSFILMZ on a specific account and generation.

The short answer: yes, but it's a first cut

Yes, Dreamina produced 180 seconds of narrative in a single generation. The video contains three scenes, dialogue, recurring characters, a title, cuts, and dramatic progression.

No, this does not mean that a solitary text produced a final film. The order includes four character sheets, names linked to references, a pre-planned structure and a prompt with thousands of characters. After rendering, the creator identifies problems that would require rewriting, new generation, editing, sound or correction.

The most useful formulation is this: Long Video can generate a three-minute narrative first cut from a multimodal package. It still doesn't eliminate storyboarding, continuity, direction, editing or quality control.

What is Long Video beta

The official Dreamina page describes up to 30 seconds in standard mode and a beta extension of 5 to 180 seconds. It also features up to 50 multimodal inputs, including text, images, video, audio, scripts, character setups and storyboards.

Long Video should not be confused with a continuous three-minute shot. The test output performs internal assembly, changes scenes, creates inserts and introduces a title. Continuity is narrative and visual, not that of a single uncut take.

The official page uses stable production and continuity language. Creator testing shows that this promise needs nuance: identity holds up relatively well, but geography and microactions remain unstable.

Setup: four character sheets

JSFILMZ starts in the Create area, opens AI Video and loads four character reference sheets. Then choose Seedance 2.5 and Long Video.

In the prompt, each name is associated with its respective reference. The process allows words like Ray, Ruth and Gil to carry a visual identity rather than relying solely on textual description.

The official limit of 50 references does not mean that more is always better. In this test, four well-defined references are enough to support the cast. Each additional input must have a clear function and not contradict the others.

Why was the prompt translated into Chinese

The interface shown accepts up to 5,000 characters. The creator's English master had approximately 7,000. After translation and compression into Chinese, the sent version is 3,721 characters long.

This frees up space because a translation can represent the same idea with fewer characters. However, JSFILMZ itself warns that the translation is not one-to-one. A spatial relationship, emotional nuance, or indirect object can change without being obvious.

The safest practice is to keep both versions and a glossary with proper names, places, props and essential phrases. Back-translate critical points and check who performs each action before spending credits.

How the Creator Spent the Character Budget

The prompt focuses mainly on dialogue and location names. With little space, the creator reduces extensive descriptions of lighting, environment and art direction.

At the end he includes negatives that he usually uses in short films made with Seedance. The intention is to reserve the majority of the text for information that moves the story and use the character sheets to carry identity.

The choice worked for narrative coverage, but it comes at a price. Minor actions are simplified, blocking is left ambiguous and location can be reinterpreted. A long prompt needs to explicitly decide what can never be inferred by the model.

Configuration, credits and render time

The test uses 16:9 format, duration of 180 seconds and 480p resolution. The interface shows 20,000 credits for this option and 40,000 for 720p.

JSFILMZ chooses 480p because it is the first attempt and wants to understand prompting, continuity and upscale quality before duplicating the credits. Rendering ends after approximately 25 to 30 minutes.

Later, it estimates around 20 to 23 dollars for the 480p render and approximately 50 dollars for three minutes in 720p. These values ​​reflect the account and conversion used by the creator. The official pages consulted do not publish a universal equivalence between credits and dollars, and warn that access and credits may vary by region and account.

Why 480p is the right decision for the first test

A low-resolution version answers the most expensive question before finishing: is the story in the right order? If the photo is shown to the wrong person or the geography changes, more pixels won't fix the problem.

The creator finds the skin and identity surprisingly good at 480p. Still, a YouTube video and an interface capture are no substitute for inspecting the original file, bitrate and upscale.

Validate structure in 480p, rewrite the weak points and only then pay 720p. If there is localized editing or chunk regeneration on your account, use it before repeating the full 180 seconds.

What does the three minutes contain

The first scene takes place in a rainy diner. Ray sees a recording made in the car and the film is titled Tide Keeps. The opening takes around 40 seconds, longer than the creator considers necessary after seeing the result.

The second returns to the same diner. Ruth delivers information about a disappearance, a vessel, eleven men, a file, a manifest and clues linked to a church. It is the section with the highest density of dialogue, props and inserts.

The third moves the investigation to the ferry, with harbor master Gil, a ticket and confirmation that a car was not sent back. The church should appear clearly beyond the pier, but it is almost hidden.

What Seedance 2.5 Got Right

The model supports several narrative beats without completely losing the cast. For 180 seconds, four characters and many lines, this is the main achievement of the test.

The title is simple and functional. The model creates inserts of the eleven missing people and investigation material without specific visual references for each piece. It also keeps the characters busy with small gestures, preventing the entire dialogue from feeling like immobile reading.

The speech is understandable and the noir atmosphere remains recognizable between scenes. The result does not prove repeatability, but it shows that a long generation can already be used for prediction, rhythm reading and editorial discussion.

Flaw 1: the geography of the diner is reversed

When the camera returns to the diner, the orientation appears mirrored. Ruth's window and position no longer correspond to the axis established in the first segment.

This is a problem of spatial logic, not just identity. People can have the same face and clothes, but the scene remains wrong if they switch sides, the exterior changes or the camera crosses the axis without intention.

To reduce error, describe a minimal plan: window to the north, Ray sitting on the side of the wall, Ruth opposite, door behind Ray. Add a layout image or storyboard and repeat the same screen direction in each scene.

Failure 2: the photograph is shown to the wrong person

The prompt asks Ruth to slide a photograph across the table. The output turns the gesture into a presentation directed at the camera instead of Ray.

The difference is small in language, but big in staging. Showing a photograph does not automatically define the recipient, trajectory, image orientation or character's gaze.

Write subject, object, recipient and final state: Ruth holds the photograph by the edge, turns it so Ray can see it, slides it to his right hand, drops the object and observes his reaction. Don't look at the camera.

Failure 3: omitted or approximated actions

The model ignores the action of removing a rubber band from the file. In another passage, a character should fold a napkin, but only manipulates something with his hands.

The result retains the general intention, busy hands and nervousness, but loses the precise action. This confirms an implicit hierarchy: dialogue and big beats survive better than micro-choreographies with props.

If an action is necessary for continuity or a narrative clue, turn it into an isolated, observable beat. If it's just texture, accept an approximation or replace it in the assembly.

Flaw 4: Time is not distributed as an editor would

The first part takes about 40 seconds before the title. Some lines have a slower pace than intended, while information at the end needs to fit into the remaining time.

A prompt with scenes in order does not guarantee proportional duration. The model may spend time on atmosphere, pauses, or movements that appear cinematic but compress later beats.

Include approximate ranges and a limit per beat. For 180 seconds, a structure of four 40-second blocks with 20 seconds reserved for transitions and title is more manageable than four scenes with no time budget.

What a prompt doesn't mean in this context

It does not mean absence of preproduction. The creator already has an 18-minute short, a cast, character sheets, relationships, locations and a text master.

It does not mean a single plan. The system creates cuts, scenes and inserts internally. It also doesn't mean a single attempt for the complete project, because the creator plans to fix the master and test again.

Finally, it doesn't mean fixed cost for three minutes. An approved release may require low-res attempts, regenerations, 720p, upscale, editing, audio, subtitles, and delivery.

A repeatable workflow in seven steps

1. Write a three-minute synopsis and limit it to three or four scenes. 2. Make character sheets with unique names. 3. Create a simple floor plan of each location. 4. Write a master in English with beats and tempo.

5. Translate only if it has a measurable advantage and back-translate critical elements. 6. Generate in 480p and evaluate with a grid. 7. Only fix blocks that failed before higher resolution generation.

The goal of the first attempt is to find out where the system interprets freely. The second version must convert each misinterpretation into an instruction, reference or simplification.

How to evaluate a 180 second output

Narrative: do the beats appear in the correct order? Identity: face, age, clothes and voice are maintained? Geography: sides, axes, entrances and windows remain coherent? Props: do objects pass to the right person and maintain shape?

Pacing: Does each scene use the allotted tempo? Audio: Are dialogue, ambience and pauses understandable? Image: is there flicker, drift or insufficient detail? Security: Do unauthorized brands, texts, faces or content appear?

Mark each issue with timecode and severity. A cosmetic error can be corrected with powders. An error in history, identity or geography can force regeneration.

When to use 30 seconds and when to use 180

Use 30 seconds when you need precise direction, a paid campaign, readable product, complex physical action, or multiple controlled attempts. The cost of repeating is lower and review is faster.

Use 180 seconds for previs, narrative animatic, proof of concept, dialogue rehearsal and continuity exploration. The value is in understanding the story as a whole, even if some shots need replacing.

A hybrid strategy is to generate the long draft, identify the best blocks and reconstruct only the critical shots in standard mode or with additional references.

How to calculate the real cost of a short trip

Useful value is not price per generation. It is approved cost per minute. Add 480p drafts, 720p versions, replacement clips, upscale, storage, editing, music, voices, subtitles, and human hours.

The creator's first render is just three minutes of a planned 18-minute short. Multiplying the price by six assumes that all blocks work the first time and that none need correction, a hypothesis that the review itself contradicts.

Record credits, queue duration, render time, number of references, prompt version, clips used and rejection reason. This history transforms curiosity into production budget.

480p, 720p and what upscale doesn't solve

Upscale can improve edges, texture and perception of detail. It does not recover a photo shown to the wrong person, a missing church, a mirrored room or an omitted action.

Compare the first and last frames of each shot, check the face, hands, text and architectural lines. Only upscale after accepting content, timing and continuity.

The Seedance 2.5 overview page advertises output up to 4K, but this specific Long Video test uses 480p and 720p options. Don't treat the general 4K claim as proof that the 180-second beta delivers 4K to your account.

Rights, consent and security

Use your own or licensed characters, voices, images, music and scripts. A consistent character sheet can come close to a real person even without mentioning their name.

Please confirm Dreamina's current terms for uploads, training, retention, outputs and commercial use. Model and credit availability may vary by region and account.

Save the prompt, references, proof of license and exported files. For client work, perform human review before publication and legal validation appropriate to the campaign territory.

Verdict: a leap to previs, not yet a final assembly

The test shows an important change: it is now possible to request a three-minute narrative block and receive something that can be viewed, discussed and rewritten as a sequence. This brings video generation closer to a cinematographic development process.

At the same time, the errors are concentrated precisely where a feature film cannot improvise: geography, recipient of objects, timing and visual information necessary for the story.

Seedance 2.5 Long Video deserves to be tested as a first-cut machine. The professional workflow continues to require multimodal preparation, timecode evaluation, replacement shots and human assembly.