Vibeleaderboard
ByteDance Seed

ai lab

Also indexed as seed · seed-bytedance

ByteDance Seed

ByteDance Seed matters because few model labs combine a comparably broad research portfolio with an existing global content, creation, and distribution infrastructure. That combination can accelerate product feedback and adoption, but the MPA dispute and independent tests of ByteDance visual models also show how quickly model capability can outrun copyright, likeness, and provenance controls when distribution is already available at scale.2,6,8

Official site ↗Models in the Atlas →

Profile

Overview

A foundation-model lab inside ByteDance

ByteDance Seed is the foundation-model research organization ByteDance established in 2023. It sits inside the larger company founded in 2012 by a team led by Zhang Yiming and Liang Rubo, but the lab is a distinct dossier subject rather than a synonym for every ByteDance product. Seed publishes research and builds models across language, infrastructure, vision, speech, multimodal interaction, world models, science, robotics, and responsible AI.1,2

General models and product distribution

Seed's general-model line progressed from Seed1.x to Seed2.0, with model cards emphasizing reasoning, long-context work, tool use, computer interaction, and economically valuable tasks. The organization connects those models to ByteDance applications and cloud services, giving it a route from research evaluation to very large user populations. Claims about hundreds of millions of users come from Seed's own model card and should be treated as company-reported reach rather than an independent usage audit.5

Seedance and Seedream

The vision group created two prominent media families. Seedance 1.0 used multimodal positional encoding, multi-shot training, video-specific reinforcement learning from human feedback, and distillation for faster generation. Seedream 4.0 unified text-to-image generation, image editing, and multi-image composition through multimodal post-training. These papers document substantial engineering contributions without establishing that ByteDance was first to invent video diffusion, multimodal image editing, or reinforcement learning for media generation.3,4

A research agenda beyond media

Seed also operates beyond media generation. Its public research index includes work on audio-visual conversation, real-world agent learning, three-dimensional assets, protein structure and binder design, theorem proving, robotics, and transformer architecture. This range supports the description of Seed as a broad industrial research organization, although the public site gives projects and papers more prominence than a stable organization chart.2

Deployment governance and originality

The lab's deployment history includes a serious governance episode. The Motion Picture Association sent ByteDance a cease-and-desist letter concerning Seedance 2.0 and Seedream 5.0 Lite. Separately, independent testing of earlier ByteDance visual models showed that they could reproduce recognizable copyrighted characters and real-person likenesses. ByteDance said it would strengthen safeguards, and in August 2026 announced an agreement with the Motion Picture Association covering filters, provenance measures, and continued cooperation for Seedance and Seedream. Reuters also reported that Zhang Yiming told Seed staff not to use rival-model distillation as a shortcut, even at a short-term competitive cost.6,8,7

Reporting and context

Notable contributions

  1. 01Native multi-shot video generationSeedance 1.0 trained for multi-shot output with consistent subjects, multimodal positional encoding, video-specific preference optimization, and a documented acceleration stack.3
  2. 02Unified image generation and editingSeedream 4.0 combined text-to-image synthesis, instruction-based editing, and multi-image composition in one multimodally post-trained system.4
  3. 03General models evaluated on real-world tasksSeed2.0 documented language, reasoning, tool-use, computer-use, long-context, and multimodal evaluation within one general-model program tied to production use cases.5
  4. 04Broad industrial AI research portfolioSeed maintains active programs spanning models, infrastructure, media generation, robotics, science, and responsible AI, with papers and project pages collected in a public research index.2