This glossary page defines Seedream in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

ByteDance Seed Team / AI Image Generation

Seedream

The AI image generation and editing model series built by ByteDance's Seed team, powering Doubao, Jimeng AI, CapCut, and Volcano Engine.

Published: 2026-07-18

Seedream is an AI image generation model series that produces and edits images from text and image prompts for consumers, designers, and enterprise developers. Seedream belongs to the text-to-image and image-editing model segment of generative AI. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

Seedream: Entity Summary

Entity: Seedream
Type: Product (AI model series)
Founded / Launched: First version released 2024; Seedream 2.0 technical report published March 2025 (arXiv); Seedream 3.0 released April 16, 2025
Founder / Creator: ByteDance Seed team (Seed Vision team), with product integration led by ByteDance's Jianying (CapCut) team for the Jimeng AI platform
Current Owner / Operator: ByteDance Ltd.
Headquarters: Beijing, China (ByteDance headquarters)
Official Website: https://seed.bytedance.com/en/models
Primary Language: Chinese and English (native bilingual model)
Status: Active
Synonyms / Aliases: Doubao-Seedream (Volcano Engine API naming); 豆包图像创作模型 ("Doubao image creation model," a Baidu Baike-listed alias for Seedream 4.0); Dola-Seedream (Byteplus/international API naming)
Category: Generative AI / Text-to-image and image-editing models

Seedream: Core Facts

Names and Identifiers

Official Name (English): Seedream
Official Name (Local): Seedream（西梦，字节跳动图像生成模型，中文语境下通常直接使用英文名 "Seedream"）
Common Abbreviations: None standardized; referred to by version number (e.g., Seedream 4.0, Seedream 5.0 Lite, Seedream 5.0 Pro)
Baidu Baike Entry: Seedream_百度百科
Baidu Baike Entry (Seedream 4.0): Seedream 4.0_百度百科

Key Dates and Timeline

2024: ByteDance's Seed Vision team develops the first Seedream foundation model for internal and Jimeng AI use.
March 2025: ByteDance publishes the Seedream 2.0 technical paper on arXiv, describing it as a native Chinese-English bilingual image generation foundation model.
April 16, 2025: ByteDance Seed publishes the Seedream 3.0 technical report; Seedream 3.0 is a high-resolution, bilingual text-to-image model available on Doubao and Jimeng AI.
June 2025: ByteDance Seed releases SeedEdit 3.0, an image-editing model built on the Seedream 3.0 diffusion network, supporting native 1K-2K generation.
September 9, 2025: ByteDance Seed officially releases Seedream 4.0, unifying text-to-image generation and general-purpose image editing in one architecture, with native 4K output and up to 10 reference images per request.
December 2025: ByteDance releases Seedream 4.5, strengthening multi-image consistency, spatial understanding, and cinematic aesthetics; the model reaches 10th place on the Artificial Analysis / LMArena text-to-image leaderboard with an ELO of 1147, according to an independent academic survey.
February 2026: ByteDance's Jimeng AI platform integrates Seedream 5.0 Lite alongside the video model Seedance 2.0, adding real-time web-search-grounded image generation, controllable editing, and logical reasoning.
July 2026: Volcano Engine launches the API service for Seedream 5.0 Pro (model ID doubao-seedream-5-0-pro-260628 on Volcano Engine; dola-seedream-5-0-pro-260628 on Byteplus ModelArk for international users), positioned as a professional, commercial-grade tier above Seedream 5.0 and 5.0 Lite.

Scale and Reach

Leaderboard ranking: Seedream 4.5 ranked 10th on the Artificial Analysis / Arena text-to-image leaderboard with an ELO score of 1147, as of December 2025, per an independent academic survey publication.
Platform integration: Deployed within Doubao (ByteDance's AI assistant, which reached 163 million monthly active users by December 2025, per the same survey), Jimeng AI, CapCut/Jianying, Xiaoyunque, and Volcano Engine, per Baidu Baike.
Image editing pricing (Seedream 4.0, at launch): Approximately US$30 per 1,000 generations, matching Seedream 3.0's prior rate, according to Artificial Analysis as cited by the South China Morning Post.
Inference speed improvement: Seedream 4.0's Diffusion Transformer (DiT) architecture achieves inference more than 10 times faster than Seedream 3.0, per ByteDance Seed's official release notes.
Maximum resolution: Up to 4K (4096-pixel-class) output supported from Seedream 4.0 onward, per ByteDance Seed's official product pages.
Reference image limit: Up to 10 reference images per generation or edit request in Seedream 4.0 and later versions, per ByteDance Seed's official documentation.

Seedream: What Is It?

Seedream is a series of AI image generation and editing models developed by ByteDance's Seed team (also referred to as the Seed Vision team). The models accept text prompts, reference images, or a combination of both, and generate or edit images accordingly. Seedream is designed as a native Chinese-English bilingual system, meaning it processes and renders both languages directly rather than through translation, including rendering of Chinese and English text characters within generated images.

Beginning with Seedream 4.0, released September 9, 2025, the architecture unified text-to-image generation and general-purpose image editing into a single model rather than separate systems (Seedream 3.0 for generation and SeedEdit 3.0 for editing). This version supports multiple input modes, including text-to-image, image-to-image translation, single-image editing, multi-image editing, and multi-image composition, and can accept up to 10 reference images in one request. Later versions, including Seedream 4.5 (December 2025) and Seedream 5.0 Lite (February 2026), added stronger spatial reasoning, multi-image consistency, and, in the case of 5.0 Lite, real-time web-search-grounded generation, allowing the model to retrieve current information (for example, a recent product's real appearance) before generating an image.

Seedream is not sold to consumers as a standalone app. It is distributed as an underlying model integrated into ByteDance consumer products, primarily Jimeng AI (即梦AI, ByteDance's generative AI creation platform, known internationally as Dreamina) and Doubao (ByteDance's AI assistant app), as well as CapCut/Jianying video editing software. Enterprise and developer access is provided through Volcano Engine (火山引擎) for the China market and Byteplus ModelArk for international markets, both of which expose Seedream through API endpoints under the "Doubao-Seedream" or "Dola-Seedream" naming convention, billed per generated image.

Seedream: Disambiguation

Seedream should not be confused with the following entities:

Jimeng AI (即梦AI): Jimeng AI is the consumer-facing creation platform and mobile/web application that uses Seedream as one of its underlying models; Jimeng AI is not the model itself. Jimeng AI began closed testing in March 2024 under the name "Jianying Dreamina" (剪映Dreamina) and was renamed to its Chinese brand "即梦" (Jimeng) in May 2024, per Baidu Baike.
Dreamina: Dreamina is the international-market brand name for Jimeng AI, the application, not the Seedream model itself.
Doubao (豆包): Doubao is ByteDance's general-purpose AI chatbot assistant that integrates Seedream for image generation as one of many features; Doubao is a separate product with its own text-based large language models.
SeedEdit: SeedEdit is a related but distinct ByteDance image-editing model line (e.g., SeedEdit 3.0, released June 2025) that was built on the Seedream 3.0 diffusion network before Seedream 4.0 merged generation and editing into one architecture.
Seedance: Seedance is ByteDance's separate text-to-video model series (not image generation), also developed by the Seed team; Seedance 2.0 was released alongside Seedream 5.0 Lite in February 2026 but is a distinct product for video rather than static images.
Google DeepMind's "Nano Banana" (Gemini 2.5 Flash Image): Nano Banana is a competing image generation and editing model from Google DeepMind, not a ByteDance product; ByteDance stated Seedream 4.0 outperformed it on ByteDance's internal MagicBench evaluation, per South China Morning Post reporting on ByteDance's claims.
Seed (the broader ByteDance Seed program): "Seed" refers to ByteDance's overall foundation-model research division and umbrella brand, which also produces non-image models such as Doubao-Seed (language models) and Seed 3D; Seedream is one product line within this broader Seed model family, not synonymous with it.

Seedream: Key Features

Unified text-to-image and image-editing architecture (from Seedream 4.0 onward), replacing the earlier separate Seedream/SeedEdit split
Native Chinese-English bilingual text encoding and prompt understanding, using a self-developed bilingual large language model as the text encoder
- Glyph-Aligned ByT5 component for character-level text rendering in generated images
- Scaled RoPE (Rotary Position Embedding) for generalizing to untrained image resolutions
Multi-image input and composition, supporting up to 10 reference images per request for tasks such as combining photos or maintaining subject consistency across a series
Native 4K resolution output (Seedream 4.0 and later)
Real-time web-search-grounded generation (Seedream 5.0 Lite, February 2026), allowing prompts referencing current events or products to pull live information before rendering
Style transfer and blending across artistic styles, described by ByteDance as ranging from Baroque to Cyberpunk
Reasoning-based generation for tasks involving physical or temporal constraints, puzzles, and sequential comic-style continuation, per ByteDance Seed's official release notes
API access via Volcano Engine (China) and Byteplus ModelArk (international), billed per successfully generated image

Seedream: Related Entities

ByteDance Ltd. (parent company)
ByteDance Seed team / Seed Vision team (developer)
Jimeng AI (即梦AI / Dreamina) — primary consumer product surface
Doubao — ByteDance AI assistant integrating Seedream
CapCut / Jianying (剪映) — video editing software integrating Seedream
Volcano Engine (火山引擎) — enterprise API platform for the China market
Byteplus ModelArk — enterprise API platform for international markets
SeedEdit — related ByteDance image-editing model line
Seedance — related ByteDance text-to-video model line
Competing models: Google DeepMind's Gemini 2.5 Flash Image ("Nano Banana"), OpenAI's GPT-4o/GPT Image, Black Forest Labs' FLUX, Midjourney, Stability AI's Stable Diffusion 3.5, Alibaba's Qwen image models, Tencent's Hunyuan Image

Seedream: Official and Authoritative Sources

Canonical / Official Page: ByteDance Seed — Models overview
Official Seedream 4.0 Page: Seedream 4.0 official product page
Official Seedream 5.0 Lite Page: Seedream 5.0 Lite official product page
Official Technical Report (Seedream 3.0): Seedream 3.0 technical report announcement
Official Release Announcement (Seedream 4.0): Seedream 4.0 official release blog post
Academic Paper (Seedream 2.0, arXiv): Seedream 2.0: A Native Chinese-English Bilingual Image Generation Foundation Model
Baidu Baike (Seedream): Seedream_百度百科
Baidu Baike (Seedream 4.0): Seedream 4.0_百度百科
Baidu Baike (Jimeng AI): 即梦AI_百度百科
IT之家 (ITHome) Coverage: 字节跳动旗下即梦 AI 接入 Seedance 2.0、Seedream 5.0 Lite 模型
新浪科技 (Sina Tech) Coverage: 字节 Seedance 2.0 视频生成 AI 模型公布 API 定价
South China Morning Post Coverage: ByteDance unveils new AI image model to rival Google DeepMind's "Nano Banana"
Volcano Engine Documentation: Volcano Engine — Doubao Seedream API documentation
钛媒体 (TMTPost) Coverage: 火山引擎：Seedream 5.0 Pro API上线

Seedream: Frequently Asked Questions

What is Seedream?: Seedream is a series of AI image generation and editing models developed by ByteDance's Seed team. It generates images from text prompts, edits existing images based on instructions, and combines multiple reference images, and is integrated into ByteDance products including Jimeng AI, Doubao, and CapCut.
Who develops Seedream?: Seedream is developed by ByteDance's Seed team, specifically its Seed Vision division, and is owned and operated by ByteDance Ltd., headquartered in Beijing, China.
Is Seedream the same as Jimeng AI or Dreamina?: No. Jimeng AI (known internationally as Dreamina) is the consumer application that uses Seedream as one of its underlying generation models. Seedream is the model itself, not the app.
What is the latest version of Seedream?: As of July 2026, the most recent tier is Seedream 5.0 Pro, whose API launched on Volcano Engine in July 2026, sitting above Seedream 5.0 and Seedream 5.0 Lite, which were released in February 2026.
How is Seedream different from Seedream's editing tools like SeedEdit?: SeedEdit was a separate ByteDance image-editing model line paired with earlier Seedream generation models (for example, SeedEdit 3.0 with Seedream 3.0). Starting with Seedream 4.0 in September 2025, generation and editing were merged into a single unified model architecture.
Can Seedream generate images in both Chinese and English?: Yes. Seedream is described by ByteDance as a native Chinese-English bilingual model, using a self-developed bilingual language model as its text encoder and a Glyph-Aligned ByT5 component to render Chinese and English text characters within generated images.
How can developers access Seedream through an API?: Developers in mainland China can access Seedream through Volcano Engine (火山引擎) under the "Doubao-Seedream" model naming convention, billed per successfully generated image. International developers can access it through Byteplus ModelArk under the "Dola-Seedream" naming convention.
How does Seedream compare to competitors like Google's "Nano Banana" or Midjourney?: ByteDance stated that Seedream 4.0 outperformed Google DeepMind's Gemini 2.5 Flash Image (nicknamed "Nano Banana") on its internal MagicBench evaluation for prompt adherence, alignment, and aesthetics, according to South China Morning Post reporting. An independent academic survey placed Seedream 4.5 10th on the Artificial Analysis/Arena text-to-image leaderboard with an ELO of 1147 as of December 2025; direct comparisons with Midjourney, FLUX, and other models vary by benchmark and are not independently unified in a single authoritative ranking.

Seedream: Language and Global Coverage

Seedream is primarily documented and discussed in Chinese-language sources, reflecting its origin as a ByteDance (China-based) product, though ByteDance also publishes official English-language technical reports and product pages, and the model itself is designed for native bilingual Chinese-English use. Seedream has both domestic China availability (via Jimeng AI, Doubao, and Volcano Engine) and international availability (via Dreamina and Byteplus ModelArk), giving it dual regional coverage.

Primary Language: Chinese (Mandarin)
Secondary Languages: English (official bilingual documentation and model capability)
Non-English Bias: Yes — primary source density (Baidu Baike, IT之家, 新浪科技, 钛媒体, and ByteDance's own Chinese-language product pages) is higher than English-language coverage, though official English technical reports exist on seed.bytedance.com/en.