ByteDance API: Seedance, Seedream and OmniHuman in One Key
Every ByteDance model on each::labs runs through one prediction API. Here's which Seedance, Seedream and OmniHuman model to pick, a real first request, polling and webhooks, and a four-step chain from one prompt to a finished explainer.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

ByteDance ships generative models faster than most teams can wire them up. A new Seedance, a new Seedream, a talking-avatar model, a voice model, all in the space of a few months. The hard part of using the ByteDance API isn't any single model. It's that a real product needs three or four of them talking to each other, and every extra integration is another auth scheme, another response format and another thing to break at 2 a.m.
So here's the short version. On each::labs, every ByteDance model on the ByteDance provider page runs through one prediction API with one key. Same endpoint, same polling, same webhooks, whether you're calling Seedance for video, Seedream for images or OmniHuman for a talking presenter. This guide walks through which model does which job, a real first request, and a chained workflow that turns one prompt into a finished clip.
What the ByteDance API Covers on each::labs
ByteDance's research arm, ByteDance Seed, builds the families you'll see here. On each::labs they fall into five groups, and knowing the groups matters more than memorizing slugs.
Seedance is video. The current flagship is Seedance 2.5 Text to Video, with siblings for image-to-video, reference-to-video and a Draft Completion endpoint. Seedance 2.0 is still live too, in a standard tier, a Fast tier and a Mini tier, each with text, image and reference inputs. Older Seedance 1.5 Pro and 1.0 Pro endpoints remain available for projects that already depend on them.
Seedream is images. Seedream v5 comes as Pro (text-to-image and edit), Lite (text-to-image and edit) and a Layer Decomposition model that splits a finished image into separate transparent layers. Seedream 4.5 and 4.0 are there as well.
OmniHuman turns a portrait plus an audio track into a talking, moving person. You get OmniHuman 2.0, OmniHuman 1.5 and the original OmniHuman. DreamActor v2 takes a different route to animation: it copies motion, expressions and lip movement from a driving video onto a character image. And Seed Audio 1.0 handles speech, with up to 120 seconds of output per request.
That's a full production stack from one vendor. Stills, motion, voice, presenters. The question is which piece to reach for first.

Pick the Model by the Job, Not the Version Number
Newer isn't automatically right. The right ByteDance model depends on what the clip is for and how many times you'll regenerate it before you're happy.
For a finished hero video, start with Seedance 2.5. Its schema accepts any duration from 4 to 30 seconds (or auto), resolutions up to 1080p, seven aspect ratios including 21:9, and synchronized audio that's on by default. If you need a long single take, that 30-second ceiling is the reason to pick it. We covered what that changes for storytelling in Seedance 2.5: Thirty Seconds in One Take.
For volume work (ad variants, social cuts, A/B tests) Seedance 2.0 Fast and Seedance 2.0 Mini are the better fit. Both top out at 15 seconds and 720p, and both keep the audio track. Use them while you're still deciding what the shot is, and graduate to 2.5 once you know.
For stills, Seedream v5 Pro is the precise one: preset aspect ratios at 1K, 1.5K or 2K, or a custom width and height. Seedream v5 Lite renders at 2K or 3K and can return up to six images in a single request through num_images, which makes it the batch tool. If your designers need editable files rather than flat pictures, run the result through Layer Decomposition, which returns up to 16 PNG layers plus a base image. The Seedream 5.0 Pro layered image post goes deeper on that.
For people on camera, OmniHuman 2.0 is the default. It accepts audio from 4 to 300 seconds, so a five-minute explainer is one request, not a stitching job. Reach for DreamActor v2 instead when you already have a performance on video and want a different character to perform it.

Make Your First ByteDance API Request
Every model uses the same create call. You send the model slug, a version (0.0.1 for all the ByteDance entries) and an input object whose fields come straight from the model's request schema. Each model page on each::labs has an API tab with the full schema, enums and defaults.
Here's a Seedream v5 Lite request. It's a good first call because it's quick and the response is easy to read.
curl -X POST https://api.eachlabs.ai/v1/prediction/ \
-H "X-API-Key: $EACHLABS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "bytedance-seedream-v5-lite-text-to-image",
"version": "0.0.1",
"input": {
"prompt": "Studio product photo of a matte ceramic coffee cup on a walnut table, soft window light from the left, shallow depth of field",
"aspect_ratio": "16:9",
"resolution": "2K"
}
}'The API answers right away with a confirmation and an ID, not the image:
{
"status": "success",
"message": "Prediction created successfully",
"predictionID": "01e83ced-f93c-4527-9b1a-b39d62c6f30d"
}That "status": "success" means the job was accepted. The picture comes next. We ran exactly this request while writing the post, and the finished prediction came back in about 25 seconds with a 2848 by 1600 JPEG: a pale ceramic mug on dark walnut, the light falling in from the left just as the prompt asked.
Poll for the Result, or Let a Webhook Call You
To fetch the output, call the same prediction ID with a GET:
curl https://api.eachlabs.ai/v1/prediction/01e83ced-f93c-4527-9b1a-b39d62c6f30d \
-H "X-API-Key: $EACHLABS_API_KEY"Repeat until status reads success, then read output. For our Seedream Lite run, output was a list holding one image URL on the each::labs CDN, because Lite can return several images from one request. The response also echoes your input and includes a metrics.predict_time field, which is handy for logging how long each model takes in your pipeline. A poll every few seconds is plenty for images. For a 30-second Seedance clip, poll less often.
Polling is fine for scripts. In production, use a webhook. Add webhook_url to the create request and each::labs calls your endpoint when the prediction finishes, so your server isn't sitting in a loop. Add webhook_secret as well and the callback is signed, which lets you reject anything that didn't come from each::labs.
{
"model": "bytedance-seedance-2-5-text-to-video",
"version": "0.0.1",
"webhook_url": "https://your-app.example.com/hooks/eachlabs",
"webhook_secret": "your-signing-secret",
"input": {
"prompt": "Slow dolly in on a lighthouse at dusk, waves breaking below, a keeper's voice says \"Lamp's lit. Bring them home.\"",
"duration": "12",
"resolution": "1080p",
"aspect_ratio": "16:9",
"generate_audio": true
}
}Notice the spoken line sits in double quotes. Seedance 2.5's schema recommends that for the best speech results, and it's the single easiest thing to get wrong in a video prompt. For more on structuring those prompts, see How to Write Better Seedance 2.5 Prompts.

Draft First, Finish Later With Seedance 2.5
This is the feature most people skip, and it's the one that changes how you iterate. All three Seedance 2.5 generators (text, image and reference to video) accept "draft": true. The model then returns a quick 480p preview plus a draft_id. If you like what you see, send that ID to Seedance 2.5 Draft Completion within seven days and it renders the final 1080p video.
{
"model": "bytedance-seedance-2-5-draft-completion",
"version": "0.0.1",
"input": { "draft_id": "the-draft-id-from-your-preview" }
}Why does this matter? Because the mistake that hurts in AI video isn't a bad render. It's a good render of the wrong idea. Drafting lets you review ten directions at preview speed and only finish the two that earned it. Build your pipeline around it from day one and your review loop gets much shorter.
Chain Seedream, Seedance and OmniHuman Into One Workflow
Single calls are easy. The real payoff of a single ByteDance API key is chaining, where one model's output URL becomes the next model's input. Here's a four-step build for a short product explainer with a presenter and b-roll.
Step 1: Generate the presenter and the product still
Call Seedream v5 Pro Text to Image twice. One prompt for a front-facing presenter portrait at 3:4, one for the product on its set at 16:9. Ask for "resolution": "2K" on both so the downstream models get clean detail.
Front-facing portrait of a friendly woman in her thirties in a linen shirt, standing in a bright ceramics studio, soft daylight, looking straight at the camera, neutral expression, mouth closed
A neutral, closed-mouth starting face gives the talking-head model the most room to animate speech naturally.
Step 2: Write the voice track
Send the script to Seed Audio 1.0. Its schema lets you shape delivery with speech_rate, pitch_rate and loudness_rate, and choose mp3 or wav through format. You can steer the voice with a speaker ID, up to three reference audio clips, or a single reference image, but only one of those three at a time.
Step 3: Animate the presenter
Pass the Step 1 portrait URL and the Step 2 audio URL to OmniHuman 2.0:
{
"model": "byteplus-omnihuman-2-0",
"version": "0.0.1",
"input": {
"image_urls": ["https://cdn-us.eachlabs.ai/uploads/your-portrait.jpg"],
"audio_url": "https://cdn-us.eachlabs.ai/uploads/your-voiceover.mp3",
"prompt": "She gestures toward the shelf on her right while speaking, slow push in"
}
}The prompt is optional, but it's where you direct movement and camera. Leave it empty and you get a competent talking head. Fill it and you get a performance. Our OmniHuman 2.0 deep dive shows what that looks like over a full five minutes.
Step 4: Turn the product still into b-roll
Feed the product image to Seedance 2.5 Image to Video with a motion prompt, and add an end_image_url if you want the shot to land on a specific final frame. Keep this step to product and scene shots: the schema notes that Seedance 2.5 doesn't accept images containing real human faces, which is exactly why the presenter goes through OmniHuman instead. Set generate_audio to false here if the voice track from Step 2 will play over the b-roll.
Four calls, one key, one response format. Wire them with webhooks and the whole explainer assembles itself while you're reviewing something else. If you'd rather not write the glue code, each::labs AI Flows lets you connect the same models visually and run the chain as one unit.

Read the Schema Before You Guess a Field
The fastest way to waste a day with any ByteDance API integration is assuming two models share field names. They often don't. Seedance 2.5 Image to Video takes a single image_url. OmniHuman 2.0 takes image_urls as an array. Layer Decomposition wants image, and so does Seedream v5 Pro Edit, which accepts 1 to 10 references for edits and multi-image blends. Seedream v5 Pro sizes output with resolution plus aspect_ratio, while Seedream 4.5 uses an image_size enum such as landscape_16_9 or auto_4K.
Durations are strings in the Seedance schemas ("12", not 12), and the allowed values differ by model: up to 30 on Seedance 2.5, up to 15 on the 2.0 family. Seedance 2.0 Fast and Mini stop at 720p. These details are all listed on each model's API tab, and checking them once beats debugging a validation error later. One more habit worth adopting: when a model accepts a language list, check it. Seedance 2.5 supports English plus nine other languages for prompts and speech, a wider set than Seedance 2.0.
Seedream 5.5 is also coming to each::labs; watch the ByteDance page for it.
Frequently Asked Questions
Is there one ByteDance API key for Seedance, Seedream and OmniHuman?
On each::labs, yes. Your each::labs key goes in the X-API-Key header and works for every ByteDance model in the catalog. You switch models by changing the model slug in the request body, nothing else about the call changes.
Which ByteDance models can I call through the API today?
Seedance 2.5 (text, image and reference to video, plus Draft Completion), Seedance 2.0 in standard, Fast and Mini tiers, Seedream v5 Pro, Lite and Layer Decomposition, Seedream 4.5 and 4.0, OmniHuman 2.0, 1.5 and the original, DreamActor v2, and Seed Audio 1.0. Older Seedance 1.x endpoints are live too. The ByteDance provider page always has the current list.
How do I know when a ByteDance video is finished?
Two options. Poll GET /v1/prediction/<id> until status is success and read output, or pass a webhook_url when you create the prediction and let each::labs call you. Webhooks are the better choice for long Seedance renders, and adding webhook_secret signs the callback.
Can I preview a Seedance 2.5 video before rendering it at 1080p?
That's what draft is for. Set it to true on any Seedance 2.5 generator, review the 480p result, then send its draft_id to Draft Completion within seven days for the 1080p version.



