Seedance and Seedream: ByteDance's video and image models

The short answer: Seedance and Seedream are two model families from Seed, ByteDance's AI research team. Seedance generates video from text, images and references, and its latest versions create the sound together with the picture. Seedream generates images from text and edits existing images based on written instructions. In short: if you want a still frame, Seedream; if you want motion and sound, Seedance. This guide covers both families' versions, the capabilities listed in official sources and how to use the two together.

Who develops Seedance and Seedream?

Both models come from ByteDance's Seed team. The team introduces its models on its own site, publishes technical reports on arXiv and offers the models in ByteDance's own products and cloud services. For example, according to the technical reports, Seedream 4.0 and Seedance 1.5 pro are accessible through Volcano Engine, ByteDance's cloud platform. Seedance 2.5, announced in July 2026, launched in apps such as Jimeng AI and Doubao Pro, with API access announced to come via BytePlus ModelArk.

The models are also offered on many third-party platforms. Length, resolution and pricing options on those platforms can differ from one platform to the next; this guide uses only the information the Seed team itself has published.

The versions in brief: which model came out when?

Model	Announced	Highlights in official sources
Seedream 3.0	April 2025	Chinese-English bilingual; fine-grained typography; native resolution up to 2K
Seedance 1.0	June 2025	Multi-shot video from text and images; 1080p
Seedream 4.0	September 2025	Generation, editing and multi-image composition in one model; native 1K-4K
Seedance 1.5 pro	December 2025	Sound and picture together; multilingual lip-sync
Seedance 2.0	February 2026	Text, image, audio and video input; 15-second multi-shot video with sound
Seedream 5.0 Pro	July 2026	Infographics, interactive precision editing, text in more than ten languages
Seedance 2.5	July 2026	30 seconds in a single pass; multi-round extension; broad reference support

What is Seedance, and what does it do?

Seedance is the Seed team's video generation model. Each version has improved both quality and the variety of inputs.

Seedance 1.0: multi-shot video

The first version, introduced in June 2025, learned text-to-video and image-to-video in the same model and natively supported narrative videos made of several consistent shots. According to its official page it generates 1080p video. The technical report gives a concrete speed figure: a 5-second video at 1080p is generated in about 41.4 seconds on an NVIDIA L20.

Seedance 1.5 pro: sound and picture together

Released in December 2025, 1.5 pro was designed to generate sound and picture at the same time. The technical report highlights precise multilingual and dialect lip-sync, cinematic camera control and more coherent storytelling.

Seedance 2.0: four kinds of input

Announced on 12 February 2026, 2.0 accepts four input types together: text, images, audio and video. According to Seed, you can provide up to 9 images, 3 video clips and 3 audio clips in one go, along with natural-language instructions. The model can take composition, motion, camera movement, effects and sound from these inputs as references. On the output side it generates 15-second, multi-shot video with two-channel audio, and it supports video extension and editing.

Seedance 2.5: 30 seconds in one go

Announced on 31 July 2026, 2.5 raised the length generated in a single pass from 15 to 30 seconds and aims at multi-minute content through several rounds of extension. The reference side grew too: up to 30 images, 10 videos and 10 audio clips at once. Seed also highlights professional features such as timestamp-level editing, green-screen editing, changing the camera perspective, and control from clay-render (white-model) references.

What is Seedream, and what does it do?

Seedream is the Seed team's image generation model. Text and typography have been a strength from the early versions on, and later versions brought editing and multi-image work into the same model.

Seedream 3.0: bilingual and strong at text

According to the April 2025 technical report, Seedream 3.0 is a Chinese-English bilingual model. It improved text rendering, especially for complex Chinese characters, and offered native high resolution up to 2K.

Seedream 4.0 and 4.5: generation and editing under one roof

Introduced in September 2025, 4.0 brought text-to-image, image editing and multi-image composition into a single framework. According to the technical report, it can generate images at native high resolutions between 1K and 4K, take several reference images and produce several outputs. The examples on the official page show edits made with a single sentence: removing a person from an image, changing the text on a poster while keeping its font and alignment, colourising an old photo and repairing its scratches. The team also released Seedream 4.5 by scaling up the same model and data.

Seedream 5.0 Pro: a model that understands design

Announced on 8 July 2026, 5.0 Pro makes claims in four areas: information-dense infographics and layouts; interactive precision editing with point selection, lasso selection, sketches, colour and material replacement, layer separation and multi-image fusion; realistic textures and skin detail; and direct input and text in more than ten languages. Besides Chinese and English, Seed names French, German, Russian, Japanese, Korean, Spanish and Arabic as examples.

Seedance or Seedream: which one for which job?

What you need	Pick	Why
Poster, cover, infographic	Seedream	Strong at text and layout; 5.0 Pro focuses on infographics
Photo editing, object removal	Seedream	Generation and editing in the same model since 4.0
A product shot in different scenes	Seedream	Multi-reference and multi-output support
A short ad or social media video	Seedance	Multi-shot storytelling and output with sound
Animating an image	Seedance	Image-to-video and image reference support
A scene with dialogue	Seedance (1.5 pro and later)	Sound and picture together, lip-sync

How do you use the two together?

The most efficient workflow is often to use the two models one after the other. In the example prompts Seed has published for Seedance, the character, the setting, the props and the storyboard are also passed to the video model as separate reference images.

1. Design the character and the setting with Seedream: generate clear front and side images of the main character and a wide shot of the setting, separately.
2. Lock consistency by editing: fix details such as outfit colour or hair with written edits, and use the same words in every image.
3. Draw a storyboard if needed: an image showing the shots in order guides the video model on camera and framing.
4. Give them to Seedance as references: state clearly in the prompt what each image stands for.
5. Extend and edit: if you like the first clip, continue the story with the extension and editing features.

What should you watch for when writing prompts?

For Seedream

Seedream's official examples use prompts that say concretely and completely what is wanted. For images with text, giving the text in quotation marks and clearly naming what should be kept (colour, font, alignment) works well:

Change the text “Summer Concert” on the poster to “Autumn Concert” and the date to “12 October”. Keep the colours, the font and the alignment exactly the same.

Create a simple infographic in a grid layout titled “Seven steps to brewing tea”. Each step gets a small illustration and a short caption; cream background, green and brown tones.

For Seedance

Seed's example prompts highlight two habits: referring to references by number (“the character from image 1”) and splitting long scenes into time ranges. Camera movement, light and sound also belong in the same prompt:

The character is from image 1, the setting from image 2. 0-5 seconds: the camera follows the character at shoulder height as they walk through the cafe door. 5-10 seconds: the character sits by the window and the camera slowly moves in on their face. Warm morning light, the clink of cups and a soft murmur of people.

Turkish is not named among the languages Seedream 5.0 Pro supports, and official sources for Seedance don't mention it separately either. For important work, writing the prompt in English and giving any Turkish text for the image verbatim in quotation marks gives more predictable results.

What are their limits?

• Complex motion: in the Seedance 2.5 announcement, Seed acknowledges there is still room to improve the physical plausibility of complex motion and the stability of scenes where several subjects interact.
• Fine text and pixel-level consistency: the Seedream 5.0 Pro announcement notes room for improvement in finer-grained text rendering and pixel-level editing consistency.
• Comparisons rely on internal tests: some of the comparison charts on the official pages come from Seed's own benchmarks (SeedVideoBench, MagicBench); don't read them as independent tests.
• Differences in access: new versions can open first in ByteDance's own apps, with API and third-party access coming later.
• Real people: if you use a real person's image or voice as a reference, get their permission, and don't show anyone doing something they didn't do.

Common mistakes

• Expecting a poster from a video model: if you need an image with dense text, prepare it with Seedream first.
• Not giving references a role: uploading several images without saying which is the character and which is the setting leads to inconsistent results.
• Describing a long video in one sentence: for a 15- or 30-second video, split the events into time ranges.
• Not describing the sound: with models that generate audio, if you don't write the ambience, music and speech, the result is left to guesswork.
• Mistaking platform limits for model limits: the app you use may restrict length or resolution; the official upper limits may not be available everywhere.

Summary

• Seedance is a video model and Seedream an image model; both come from ByteDance's Seed team.
• Seedance 2.0 makes 15-second video with sound from four kinds of input; 2.5 makes 30 seconds in a single pass.
• Seedream has combined generation and editing since 4.0; 5.0 Pro focuses on infographics and precise editing.
• Choose Seedream for still frames and text, Seedance for motion and sound.
• The best results often come from using both: design with Seedream, animate with Seedance.
• Give references a role, split long scenes into time ranges and get permission for real people.