Emu (model)
Emu is a family of generative image and video foundation models developed by Meta AI, first introduced in 2023 for photorealistic image generation and later extended into variants for editing, video generation, and multimodal…
Definition
Emu is a family of generative image and video foundation models developed by Meta AI, first introduced in 2023 for photorealistic image generation and later extended into variants for editing, video generation, and multimodal understanding. Emu underpins several of Meta's consumer creative-AI features across Instagram, Facebook, and the Meta AI assistant.
Overview
Emu (Expressive Media Universe) is Meta's line of latent diffusion foundation models for visual generation. The original Emu model, described in a 2023 Meta AI paper, was trained on a large corpus of images and then fine-tuned on a small set of a few thousand carefully curated, highly aesthetic images — a technique Meta called 'quality-tuning' — to steer the model toward producing visually appealing, photorealistic outputs without changing its architecture. This approach demonstrated that supervised fine-tuning on a small, high-quality dataset could meaningfully improve output aesthetics compared to scaling pretraining data alone. Meta subsequently expanded the Emu family into specialized variants: Emu Edit, which supports instruction-based image editing (e.g., adding, removing, or localizing changes to objects) while preserving unedited regions of an image; Emu Video, which generates short video clips from text or image prompts using a factorized diffusion approach that first generates a static image and then animates it; and Emu Video Edit / related multimodal successors that combine understanding and generation in a single model. Emu models power user-facing features across Meta's app family, including AI image generation and editing stickers in Instagram and Messenger, backdrop and restyling tools, and the image-generation capabilities of the Meta AI assistant. Meta has generally kept Emu models proprietary and accessible primarily through its own products rather than releasing open weights, in contrast to Meta's more open approach with its Llama language models. Emu is significant as an example of a major consumer platform integrating diffusion-based generative AI directly into everyday social and messaging surfaces at large scale.
Key Concepts
- Latent diffusion architecture for text-to-image and text-to-video generation
- Uses 'quality-tuning' — fine-tuning on a small curated high-aesthetic dataset
- Emu Edit supports instruction-based, localized image editing
- Emu Video generates short animated clips from text or image prompts
- Deployed at scale inside Instagram, Facebook, Messenger, and Meta AI
- Proprietary model, unlike Meta's openly released Llama LLMs
- Optimized for photorealism and consumer creative-tool use cases