All posts
Video Generation API

Seedance 2.5 Flash: When to Use It and 9 Prompts to Copy

Seedance 2.5 Flash is ByteDance's fast tier of Seedance 2.5: 30-second takes with sound at 480p or 720p. Here's when to pick it over the full model, plus nine prompts to paste in, three per mode.

By Eftal Yurtseven · Nov 16, 2026 · 10 min read

Seedance 2.5 Flash: When to Use It and 9 Prompts to Copy

Nobody gets a 30-second video right on the first try. The camera move lands on take four, the dialogue on take nine, and around take twelve you notice the opening beat was wrong all along. That's normal. The waiting is what hurts, because a model that renders half a minute of picture and sound in one pass makes you wait for every attempt.

Seedance 2.5 Flash is ByteDance's answer to the waiting: the fast tier of Seedance 2.5, with the same 30-second takes, sound, references, editing and extension, and the top resolutions traded away for speed. It's live on each::labs as text-to-video, image-to-video and reference-to-video.

Flash isn't a lesser Seedance 2.5. It's the one you think with. Here's when to pick it, and nine prompts to paste straight in.

What Seedance 2.5 Flash Actually Is

BytePlus lists Seedance 2.5 Flash among its November 2026 releases on ModelArk as dreamina-seedance-2-5-flash-261115, and documents it next to the full model in its Seedance 2.5 series tutorial. It covers the full model's task list: text-to-video, first and last frames, references (image, video, audio), editing and extension. Output runs 4 to 30 seconds, or auto, with synchronized audio in the same pass.

Three things change. Flash renders 480p or 720p, while the full Seedance 2.5 on each::labs goes up to 1080p. The reference ceiling drops from 50 assets (30 images, 10 videos, 10 audio clips) to 20 (10, 5 and 5). And there's no Draft mode, the two-step preview flow BytePlus reserves for the full model.

What you get back is speed. ByteDance's own Flash page pitches it as a way to test video ideas 2x faster, with quality close to the original. That's ByteDance's number, not a published benchmark, so judge it on your own prompts.

Fast is a setting. Final is a decision.
Fast is a setting. Final is a decision.

Seedance 2.5 Flash vs Seedance 2.5: When to Reach for Flash

Don't choose on quality versus speed. Choose on what the clip is for.

Pick Flash while you're still deciding. Camera moves, beat timing, whether a line lands, which of three openings works: all of that gets decided by watching, and Flash lets you watch more takes in the same afternoon. Run those rounds at 480p. You're judging motion and story, not skin texture.

Then there's the clip whose final home is a phone. A vertical post for Reels, TikTok or Shorts, a story ad, a preview inside a mobile app: at that size a 720p take from Flash is finished work, not a sketch. Long takes with sound belong here too. The 30-second ceiling and the audio pass match the full model's, so a half-minute short-drama scene with dialogue is a Flash job whenever the resolution fits.

Reach for the full Seedance 2.5 when the final has to be 1080p, when you need more than 10 reference images or 5 clips, or when you want to finish the exact take you previewed. That's Draft mode: switch on draft for a 480p preview, then pass its draft_id to Seedance 2.5 Draft Completion within seven days for the 1080p final.

Here's the part people miss. Flash locks the prompt, not the take. The same prompt on the full model gives you a new generation, close in intent but not frame for frame. ByteDance suggests exactly this loop: find the idea on Flash, then reuse the prompt and references on Seedance 2.5 for the high-quality version. (Against Kling? See Seedance 2.5 vs Kling 4.0.)

Pick the input before you pick the words.
Pick the input before you pick the words.

The Three Seedance 2.5 Flash Models on each::labs

All three sit on the Seedance 2.5 family page and share defaults: 720p, duration on auto, audio on.

Text to Video takes a prompt (keep it under roughly 1,000 words), an aspect ratio from 21:9 to 9:16 or auto, a duration of 4 to 30 seconds or auto, and 480p or 720p. It reads English plus nine other languages. Put spoken lines in double quotes for clean speech.

Image to Video takes a first frame and an optional last frame. The output follows your first image's aspect ratio, and a mismatched last frame gets center-cropped. Photos of real people work as inputs too (they're registered in BytePlus's private asset library first), so a selfie idea is fair game.

Reference to Video takes up to 10 images, 5 videos and 5 audio clips, called in the prompt as @Image1, @Video1, @Audio1. Video and audio each get 30 seconds combined, and audio alone can start a generation. Wording switches the task: add, remove or replace edits a reference video, extend or continue extends it. Videos with real faces are refused unless Seedance made them under your account in the last 30 days, which covers your own Flash output.

Text-to-Video Prompts for Seedance 2.5 Flash

Write Flash prompts the way you'd write for the full model (our guide on how to write better Seedance 2.5 prompts covers the building blocks): format, cast and world, timed beats, camera, sound. The difference is attitude. On Flash, a prompt is a hypothesis you expect to rewrite three times.

A vertical product teaser at 720p, 9:16, 15 seconds.

9:16, bright commercial style, single continuous take. 0 to 4s: macro close-up of a cold can of yuzu soda sweating on a sunlit concrete ledge. 4 to 9s: a hand pulls the tab and the camera pushes in as mist escapes. 9 to 15s: pull back to a rooftop at golden hour, friends laughing out of focus behind the can. Sound: tab crack, fizz, a light lo-fi beat. No text on screen.

A short-drama scene with dialogue. Test at 480p, 16:9, 30 seconds, and move to 720p once the performances work.

16:9, moody short-drama style, three shots. Shot 1, 0 to 10s: night, a young woman in a rain-soaked denim jacket waits outside a closed laundromat under a buzzing neon sign. Shot 2, 10 to 20s: a man in a grey hoodie watches her from a bus shelter, hesitates, then crosses the street. Shot 3, 20 to 30s: close two-shot under the neon; he holds out a folded umbrella and says, "You forgot this at my place." She looks at the umbrella, not at him: "I didn't forget it." Sound: steady rain, neon hum, no music.

A creator-style talking clip at 720p, 9:16, 20 seconds.

9:16, handheld phone footage, warm kitchen light, single take. 0 to 6s: a home cook in a mustard sweater holds up a pan at her counter and says, "Okay, the five-minute pasta trick." 6 to 14s: quick push in as she tosses spaghetti with butter and lemon, steam rising. 14 to 20s: she twirls a forkful toward the lens and says, "Told you." Sound: sizzling pan, a fork clinking, small-room ambience, no music.
Start from the frame you already trust.
Start from the frame you already trust.

Image-to-Video Prompts: Animate the Frame You Already Like

Image to video is where Flash pays off fastest: the composition is settled and you're only testing motion. Write movement, camera and sound; describing the image again invites the model to change it.

A packshot into a hero shot. Product photo as the first frame, 720p, 8 seconds.

The perfume bottle stays centered and upright. The camera slowly arcs right around it while warm morning light sweeps across the glass and throws long amber reflections over the marble. A few white petals drift past. Sound: quiet room tone, one soft piano note as the light settles.

A mascot move between two approved poses. Start pose as the first frame, end pose as the last frame, 480p while testing, 6 seconds.

The fox mascot hops off the arm of the couch, lands on the rug, spins once and finishes facing the camera with a wave, exactly as in the last frame. Camera locked off at eye level. Sound: a soft thump on landing, a whoosh on the spin, no music.

A creator photo that talks. A real photo of her at the desk as the first frame, 720p, 10 seconds.

She looks up from the laptop, smiles at the lens and says, "This whole video took me ten minutes." Gentle handheld drift; the window light stays exactly as in the photo. Sound: two keyboard taps before she looks up, a quiet room, no music.

Reference-to-Video Prompts: Ten Images Is Plenty

Ten images sounds like a step down from thirty until you list what most briefs need: a face, a product, a location. The real risk was never too few references. It's too many that disagree.

Character, product and place in one take. Three images, 9:16, 720p, 20 seconds.

@Image1 is the main character: keep her face, braided hair and oversized grey sweatshirt exactly as shown. @Image2 is the sneaker: keep its colorway. @Image3 is the skate park. Single handheld take. 0 to 7s: she laces up the sneaker from @Image2 on the edge of the bowl in @Image3. 7 to 14s: she drops in and the camera follows at board height. 14 to 20s: she stops at the lip, looks into the lens and says, "Broke them in already." Sound: wheels on concrete, wind, no music.

A clip cut to your track. One audio file and nothing else, 9:16, 720p, 15 seconds.

Use @Audio1 for the rhythm and mood of the whole clip. A paper boat sails down a rain-filled gutter through a city street at dusk. Cut on the beats of @Audio1, each shot a little closer to the boat. Final shot: the boat tips into a storm drain on the last hit and the frame goes dark.

Extending a Flash take you like. Pass the earlier output as the video reference, then set a duration or leave it on auto.

Extend @Video1. The courier keeps riding past the last market stall, turns into a quiet alley and brakes at a blue door. Keep the same camera height, evening light and street sound, and end on her knocking twice.

To edit instead, change the verb: "Replace the red umbrella in @Video1 with a yellow one, keep everything else." An edit keeps its source's length and aspect ratio, so leave both on auto.

Grow twenty. Transplant one.
Grow twenty. Transplant one.

Run a Flash-First Loop

Treat Flash as the first half of a pipeline. Start at 480p, duration on auto, and watch for one thing per round: the camera, then the timing, then the performance. Once a prompt survives two rounds without a rewrite, lock it, set the duration so clip and timestamps agree, and render at 720p. Need 1080p? Take the locked prompt and references to the full Seedance 2.5.

For a creator, that loop is an afternoon: twenty ideas tested, three shipped. For a mobile app studio, it's a product decision. A generate button is a promise about how long someone waits, and Flash is the tier built for that button: a 480p preview while the user chooses, a 720p export once they commit. Same loop every week? Build it once in each::labs AI Flows.

Calling the Seedance 2.5 Flash API

All three Flash models use the same prediction endpoint as every other text-to-video model on each::labs. Switching mode means changing the model string and inputs.

curl -X POST https://api.eachlabs.ai/v1/prediction \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "bytedance-seedance-2-5-flash-text-to-video",
    "version": "0.0.1",
    "input": {
      "prompt": "9:16, bright commercial style, single continuous take. A cold can of yuzu soda sweats on a sunlit ledge, a hand pulls the tab, mist escapes. Sound: tab crack, fizz, light lo-fi beat.",
      "aspect_ratio": "9:16",
      "duration": "15",
      "resolution": "480p",
      "generate_audio": true
    },
    "end_user_id": "usr_4f9a2c"
  }'

Duration is a string ("15" or "auto"), not a number. Every call also sends end_user_id at the top level, next to input: a stable, pseudonymous ID for the person behind the request (your internal user ID or a hash of it, never an email), so swap the placeholder for your own. For image to video, use bytedance-seedance-2-5-flash-image-to-video with image_url and an optional end_image_url. For references, use bytedance-seedance-2-5-flash-reference-to-video with image_urls, video_urls and audio_urls. The call returns a prediction ID; poll https://api.eachlabs.ai/v1/prediction/{id} until the status reads success, then take the video URL from the output.

The fastest way to know whether an idea works is still to watch it. Flash lets you watch more of them before lunch. Skim our Seedance 2.5 overview, then open Seedance 2.5 Flash and paste in the first prompt.

Frequently Asked Questions

What's the difference between Seedance 2.5 Flash and Seedance 2.5?

One series, two jobs. Both make 4 to 30 second clips with synchronized audio from text, frames or references, and both edit and extend video. Flash is the faster tier: it stops at 720p, takes 20 references instead of 50 and skips Draft mode. The full model goes to 1080p on each::labs.

Can Seedance 2.5 Flash make 1080p video?

Flash tops out at 720p by design. For a 1080p final, take your refined prompt to the full Seedance 2.5, or preview there with Draft mode and finish that exact take with Draft Completion.

How long can a Seedance 2.5 Flash clip be, and does it have sound?

Thirty seconds in one pass, the same ceiling as the full model. Audio comes with the picture by default: voices, effects, ambience and music. Set generate_audio to false for a silent clip.

Is there a Seedance 2.5 Flash API?

Yes: on each::labs it's the same prediction API every model uses, with one model string per input type. In ByteDance's own docs, BytePlus ModelArk lists it as dreamina-seedance-2-5-flash-261115.

Start with what you need.

Scale when you’re ready.