What is Dreamina Seedance 2.0?

Dreamina Seedance 2.0 is ByteDance's second-generation text-to-video model, built by its Seed research team. The same Seedance 2.0 engine powers ByteDance's consumer apps — CapCut and Dreamina (formerly Jimeng AI) — and is exposed for developers through API resellers like GPTProto. On this page the model runs in text-to-video mode: you send a prompt, it returns a finished clip with audio generated in the same pass.

What separates Seedance 2.0 from the 1.x line is one-pass audio-visual generation. Dialogue, sound effects, ambient noise, and music are produced together with the picture in a single generation — no separate audio step, no stitching. The model keeps a character's identity, features, and style consistent across shots, which is the part most short-form pipelines actually struggle with.

On GPTProto you reach dreamina-seedance-2-0-260128 with one model string and your existing balance, so you can route between Dreamina Seedance 2.0, Doubao Seedance 2.0, Sora 2, and Veo 3 without separate accounts or regional sign-up gates.

Specifications

Three ways to drive it: text, image & reference

Dreamina Seedance 2.0 exposes three generation endpoints under the same model id. Each takes the same core controls — aspect_ratio, duration, resolution, generate_audio, camera_fixed, seed — and differs only in what you feed it.

Text-to-video — prompt in, clip out. Endpoint: POST /api/v3/doubao/dreamina-seedance-2-0-260128/text-to-video

Image-to-video — animate a still; optionally pin a last_image to control the end frame. Endpoint: POST /api/v3/doubao/dreamina-seedance-2-0-260128/image-to-video

Reference-to-video — pass arrays of images, videos, and audios as references for character and style continuity. Endpoint: POST /api/v3/doubao/dreamina-seedance-2-0-260128/reference-to-video

All three use the same bearer auth and the same submit-and-poll flow shown below — only the input fields change.

Dreamina Seedance 2.0 vs Sora 2, Veo 3 & Doubao Seedance 2.0

All four run on GPTProto under one balance, so the real question is which to route each job to. Prices below are on-site, per run.

Honest read: Dreamina Seedance 2.0 and Doubao Seedance 2.0 are the same ByteDance generation via two brand routes — identical $0.2957/run, differing in platform defaults, not output ceiling. Across these four, Seedance is the lowest per-run option and the only one here with three distinct input modes (text, image, multimodal reference).  Veo 3 ($0.48/run) is the pick when you need 4K and tight cinematic control; Sora 2 ($0.40/run) for OpenAI-native generation. A common pattern is drafting on Seedance, then re-rendering the chosen take on a pricier model.

For half-cost drafts, dreamina-seedance-2-0-fast-260128 runs $0.2365/run.

Switching from CapCut / Dreamina to the API

If you've been generating inside CapCut or the Dreamina app, the API drops the per-app sign-up and regional rollout limits without a new SDK to learn. Point your client at https://gptproto.com/api/v3, send your key as a bearer token, and POST to the dreamina-seedance-2-0-260128 task endpoint. One balance then covers every other model, so you can A/B Seedance against Sora 2 or Veo 3 in the same job runner.

Prompt Recipes

Seedance 2.0 responds to explicit, ordered direction: style, then subject, then camera, then lighting, then motion.

1. Product hero (locked camera): cinematic; matte-black wireless earbuds on wet slate, water droplets; locked-off static shot; soft studio key light from left; slow condensation forming — set camera_fixed: true.
2. Character short (push-in): anime style; a courier cycling through a rainy neon alley at night; slow push in; reflections on the road; ambient rain and distant traffic — keep generate_audio: true for the ambient bed.
3. Reference-driven continuity: Pass two character stills in images[] and prompt same character, new shot: walking through a morning market, handheld follow — the reference mode holds identity across the cut better than a fresh text prompt.

State camera moves literally ("slow push in", "tracking shot", "locked off") and name the lighting directly. Vague prompts drift; specific ones hold.

What to expect: speed, audio & aspect ratios

Generation time. Because clips render asynchronously, plan for a wait, not an instant response. Standard shorts return quickest; long 15-second clips with multiple references take the longest — size your polling interval and any progress UI accordingly.

Audio. With generate_audio: true the model writes dialogue, effects, ambient sound, and music to match the picture in the same pass. Set it false for a silent clip when you'll mix your own track — common for music videos or branded content with a fixed audio bed.

Aspect ratio. Pick the target ratio at submit time rather than cropping afterward; the model frames the action for the canvas it's given. 16:9 is the default; the full ratio list is in the on-page table.