Seedance 2.5 Is Live on Genfire: 30-Second Multi-Shot AI Video in One Pass

What Seedance 2.5 Is

Seedance 2.5 is ByteDance's newest video generation model, and it is now the default model in Genfire's video studio. The headline is length: it renders 4 to 30 seconds of video as one continuous generation, with scene changes, camera moves, and tempo shifts inside that single pass. Nothing is stitched, so the lighting, the grade, and the performance carry across a cut instead of resetting at it.

The second change is how much you can hand it. A reference-to-video request takes up to 30 reference images, 10 reference clips, and 10 audio takes — 50 files in total — and you address each one in the prompt as @Image1, @Video1, or @Audio1.

The rest of the spec, checked against the shipped endpoint:

• Duration: 4–30 seconds, or leave it unset and the model picks a length that fits the prompt
• Resolution: 480p, 720p (default), or 1080p — no 4K
• Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, or auto
• Modes: text-to-video, image-to-video (start frame plus an optional end frame), and reference-to-video
• Audio: sound design, ambience, and lip-synced dialogue generated in the same pass as the picture, at no extra cost
• Speed tiers: none — there is one quality to choose from

The full spec sheet and a showcase of 30-second films are at /seedance-2-5. This post covers what moved from Seedance 2.0, how to prompt it, and how to run it from every Genfire surface.

What Changed From Seedance 2.0

Seedance 2.0 is still on Genfire and still worth using — it is the only Seedance that reaches 4K. But 2.5 changes the shape of the model in three ways.

Clips go from 15 seconds to 30

Seedance 2.0 generates 4 to 15 seconds. Anything longer meant separate takes cut together, and every cut was a chance for the character, the light, or the lens to drift. Seedance 2.5 doubles the ceiling to 30 seconds and lets the model change shots inside the take. A wide, a medium, and a close-up come back as one file with one grade.

References go from 15 files to 50

Seedance 2.0 accepts 9 images, 3 videos, and 3 audio clips. Seedance 2.5 accepts 30, 10, and 10. You can lock a face from six angles, a product from four, a location from three, and still have room for a motion reference and a voice take. Audio references are how a character keeps the same voice from clip to clip; they need at least one image or video reference alongside them, and it helps to say who is speaking with which take.

Speed tiers are gone, 4K stays with 2.0

Seedance 2.0 ships as Standard, Fast, and Mini. Seedance 2.5 ships as one model, and resolution is the dial instead: 480p for drafts, 720p as the default, 1080p when the clip is the finished thing. 1080p costs more per second than 720p and 480p costs less. 4K remains a Seedance 2.0 Standard exclusive — if the deliverable is a 4K master, generate it on 2.0.

How It Compares With Wan 3.0 and Veo 3.1

Several models on Genfire produce a clip with sound. The three most people weigh against each other right now are Seedance 2.5, Alibaba's Wan 3.0, and Google's Veo 3.1. These facts come straight from Genfire's model registry.

The short version: Seedance 2.5 has the deepest reference pool on the platform and the widest aspect-ratio range, including 21:9. Wan 3.0 matches it on length and audio, starts shorter at 2 seconds, defaults to 1080p, and adds a Prime tier. Veo 3.1 tops out at 8 seconds but is the one to reach for when you want first-and-last-frame interpolation or a 4K frame from Google's pipeline.

Writing Prompts for Multi-Shot Video

A 30-second single-pass clip is a different kind of prompt from a 5-second one. You are writing a shot list, not a caption. Three habits make the difference.

Set the style once, then bracket the timing

Open with one line that fixes the overall look — lens, grade, pacing, mood — then lay out the shots in order with rough time brackets. The model reads the brackets as beats rather than hard cut points, but they tell it how to spread 30 seconds across the sequence. One of the showcase films on the model page is prompted exactly this way:

[Overall style setting] A cinematic, high-quality 30-second haute couture visual blockbuster emphasizing bokeh light spots, silky motion blur, volumetric lighting, and ultra-realistic material details. [0–5s] Macro close-up: a slender hand reaches into the air, fingertips touching colorful light spots as bright as stars… [5–15s] A smooth follow shot: a woman in a French wide-brimmed straw hat and flowing white dress runs lightly through a dense flower path… [15–24s] Quiet aesthetics: a young girl reads beside a European retro gold fountain as soap bubbles drift…

If you would rather not write brackets, name the shots in sequence and separate them with "then". The model will still cut between them.

Repeat what must stay the same across cuts

Every cut is a chance for the subject to drift, so say what carries over. "Same suit, same light" at the end of a shot list is doing real work:

A lone astronaut steps out of a dust-caked rover onto a rust-red dune at golden hour. Wide crane-up revealing the canyon, then a tight push-in on her gloved hand brushing sand off a fossil, then a low hero shot as twin moons rise — same suit, same light.

Three connected shots of the same chef in a steel kitchen: a knife rocking through scallions, a wok tossing with flames leaping, then plating as steam curls toward camera. Keep her face, tattoos, and apron identical across every cut.

Be specific about anything visible in more than one shot — the face, the wardrobe, the prop, the time of day, the lens. Things that appear once can be described once.

Cite references by name and give each one a job

Address every attached reference in the prompt and say what it is for. @Image1 is a face or a product; @Video1 is a camera move or choreography to follow; @Audio1 is a voice. Pair audio with at least one image or clip and say who speaks with which take:

The woman in @Image1 walks into frame and speaks the line in @Audio1 to camera, then the shot cuts wide to the rooftop bar behind her and holds as the city lights come up.

A reference the prompt never mentions is a reference the model has no reason to bind. Audio is on by default and costs the same as turning it off, so describe the soundscape too — "kitchen ambience, no music" or "a rising synth pad under the final shot" — rather than leaving it to chance.

How to Use It on Genfire

Seedance 2.5 runs on every Genfire surface against the same model, with the same duration, resolution, and reference limits.

In the browser studio

Open the AI video generator. Seedance 2.5 is the default model, so you can type a prompt and generate immediately. Switch between text, image, and reference modes in the same panel; attach a start frame and an optional end frame for image-to-video, or drop in reference images, clips, and audio and cite them with the @Image1-style chips the studio inserts for you. Pick a resolution and a duration from 4 to 30 seconds, or leave duration on auto. The studio shows the credit cost before you generate.

Through the REST API

Call POST /v1/videos/generations with model video.seedance_2_5. Routing follows the inputs: a prompt alone is text-to-video, adding image_url makes it image-to-video (add end_image_url for the landing frame), and adding reference_image_urls, reference_video_urls, or reference_audio_urls makes it reference-to-video. Omit duration to let the model choose its own length.

curl -X POST https://api.genfire.ai/v1/videos/generations \
  -H "Authorization: Bearer $GENFIRE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "video.seedance_2_5",
    "prompt": "The woman in @Image1 walks into frame and speaks the line in @Audio1 to camera, then the shot cuts wide to the rooftop bar behind her and holds as the city lights come up.",
    "reference_image_urls": ["https://.../woman.jpg"],
    "reference_audio_urls": ["https://.../line.mp3"],
    "resolution": "1080p",
    "aspect_ratio": "16:9",
    "duration": 20
  }'

The run comes back queued; poll /v1/runs/{runId} or register a webhook, and call POST /v1/models/estimate-cost for an exact quote first. Keys and reference docs are at /developers.

Through the MCP server

Connect Claude, ChatGPT, Cursor, or any MCP client to mcp.genfire.ai and call genfire_generate_video with model: "video.seedance_2_5". The tool takes the same fields as the API — prompt, duration, aspect_ratio, resolution, image_url, reference_image_urls, reference_video_urls, reference_audio_urls — and allows 30, 10, and 10 in the pools when the model is Seedance 2.5. Setup at /mcp.

From the CLI

npm i -g @genfire/cli && genfire auth login

genfire generate video "Three connected shots of the same chef in a steel kitchen: a knife rocking through scallions, a wok tossing with flames leaping, then plating as steam curls toward camera. Keep her face, tattoos, and apron identical across every cut." \
  -m video.seedance_2_5 -d 24 -r 1080p -a 16:9 -o chef.mp4

--image and --end-image set the start and end frames; --ref-image, --ref-video, and --ref-audio fill the reference pools and accept URLs or local paths (local files upload automatically). The command waits for the run and downloads the result by default. Docs at /cli.

Pricing

Genfire is pay-as-you-go. Credit packs start at $19 for 1,000 credits, purchased credits last 12 months, and every purchase includes a commercial license; optional monthly plans start at $29 a month. Seedance 2.5 is billed per second of video — 1080p costs more per second than 720p, 480p costs less, and audio is included. The studio, the API's estimate endpoint, and genfire cost video all quote the exact number before a run starts. Current rates are on the pricing page.

Where It Fits

Seedance 2.5 is the model for a finished sequence rather than a single shot — a 30-second spot, a multi-beat social clip, a scene with a character who has to speak and stay recognisable. Keep Seedance 2.0 for 4K masters, and look at Wan 3.0 when you want a 2-second minimum, a 1080p default, or a Prime tier for speed.

Open the video studio and Seedance 2.5 is already selected. The full spec and showcase are at /seedance-2-5, and credit packs are on the pricing page.