Seedance 2.0: the complete guide

Seedance 2.0 is a diffusion-transformer video model. Like Veo 3.1, Kling 3.0 and Sora 2 it outputs native 1080p (and up to 4K), charges by the second, and clears the uncanny-valley bar most of the time. What sets it apart is not raw image quality — it is what the model accepts as input, and what it generates alongside the video.

• ▸Multimodal references: up to 9 images, 3 video clips and 3 audio files in a single generation (12 mixed inputs total). Competitors take 1–2 images; Veo 3.1 additionally takes short video refs. Nobody else comes close to 12.
• ▸Native audio in the same pass: sound effects, ambience and lip-synced dialogue are generated together with the frames. Dialogue placed in double quotes inside the prompt is lip-synced in 8+ languages. Audio costs nothing extra.
• ▸Longest usable clip in its class: 4–15 seconds per generation. A 60-second piece needs 4 chained generations instead of 8 on Veo's 8-second ceiling.
• ▸Multi-shot editing and director-level camera control (dolly, orbit, crash-zoom, speed-ramp) respond reliably to plain-text prompts.

ByteDance announced Seedance 2.5 in June 2026 (30-second clips, up to 50 references), but it is still in limited rollout. Seedance 2.0 is the version you can actually build on today, and it will remain the stable production endpoint for months.

On fal.ai, Seedance 2.0 bills per second of output video, with the effective rate derived from a token formula. The published per-second rates:

Under the hood fal computes tokens as (height × width × duration × 24) / 1024 — i.e. every output pixel of every 24 fps frame, divided by 1024. The token price is $0.014 per 1,000 tokens for 480p/720p/1080p and $0.008 per 1,000 tokens for 4K ($0.0112 per 1,000 on the fast tier). Two practical consequences:

• ▸Resolution scales cost quadratically. 1080p costs ~2.25× more than 720p for the same clip, 4K far more again despite the lower token rate. Iterate at 480p/720p, render finals high.
• ▸Audio is genuinely free — generate_audio: true does not change the token count, so there is no reason to switch it off unless you need silent footage.

Real numbers from our own lab: our 8-second 720p interior flythrough on the fast tier cost about $1.94 (~€1.80), and an 8-second 9:16 ASMR clip with full audio came in at the same rate. For comparison, the same 8 seconds on Veo 3.1 Standard is $3.20 — with a hard 8-second ceiling.

The fal.ai endpoints are bytedance/seedance-2.0/text-to-video, .../image-to-video and the fast variants. Generation is asynchronous: you submit to the queue, poll the status URL, then download the MP4. Full input schema for text-to-video:

curl -X POST https://queue.fal.run/bytedance/seedance-2.0/text-to-video \
  -H "Authorization: Key $FAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Handheld interior flythrough of a sunlit villa. \"Welcome home,\" she says warmly.",
    "resolution": "720p",
    "duration": "8",
    "aspect_ratio": "16:9",
    "generate_audio": true
  }'

Typical queue-to-download time in our runs was 1–4 minutes for an 8-second 720p clip. Budget for occasional retries: roughly 1 clip in 5 needed a second attempt to nail the brief.

• ▸Structure prompts as: shot type → subject → action → environment → lighting → camera move → audio cues. The model parses film language well.
• ▸Dialogue in double quotes gets lip-synced. Keep lines under ~15 words per clip; longer speech drifts out of sync near the end.
• ▸Name the camera move explicitly ("slow push-in", "orbit right", "FPV flythrough", "speed-ramp"). Vague verbs like "cinematic movement" produce timid pans.
• ▸For multi-scene continuity, reuse the character reference image in every generation and feed the last frame of clip N as an image reference for clip N+1.
• ▸Describe audio you want ("soft rain, distant traffic, glass clink") — the audio track follows prompt cues, not just the visuals.
• ▸Use duration: "auto" for single showcase clips — the model often picks a better pace than a forced length — but pin exact durations for anything you will edit into a sequence.

• ▸Wins: reference handling, multi-scene character continuity, dialogue work, cost-per-usable-clip at 720p, duration ceiling.
• ▸Loses to Sora 2 on raw physics realism (glass shattering, complex hand-object interaction) — though Sora 2 has a published shutdown date of September 24, 2026, so don't build on it.
• ▸Loses to Veo 3.1 on top-end cinematic polish: color science and tonal range on premium single-shot brand films still favor Google.
• ▸Kling 3.0 undercuts it per generation, but limited references mean more retries; cost per finished clip usually flips back in Seedance's favor once you need brand or character control.

Our verdict: if you are building any repeatable AI video pipeline in 2026 — ads, listings, social content, explainer clips — Seedance 2.0 on the fal.ai fast tier is the default starting point. Move individual shots to Veo 3.1 when a client is paying for the last increment of polish, and treat Kling 3.0 as the drafting tool.