Bagel

BAGEL by ByteDance-Seed is an Apache 2.0 open-source unified multimodal model designed for advanced image/text understanding, generation, editing, and navigation. It offers capabilities comparable to proprietary systems like GPT-4o and Gemini 2.0. BAGEL can be fine-tuned, distilled, and deployed anywhere, providing precise, accurate, and photorealistic outputs through its natively multimodal architecture.

How to Use

BAGEL can be used through its unified multimodal interface, accepting both image and text inputs and outputs in a mixed format. Users can engage in multi-turn conversations, generate high-fidelity images and video frames, perform image editing, apply style transfers, navigate virtual environments, and leverage its compositional and thinking modes by providing prompts and interacting with the model.

Key Features

• Unified Multimodal Model
• Image/Text Understanding
• Image/Text Generation (photorealistic images, video frames)
• Image Editing (preserves visual identities and details)
• Style Transfer
• Navigation (in diverse environments)
• Compositional Abilities (multi-turn conversations)
• Thinking Mode (enhances generation and editing through reasoning)
• Pre-training initialized from large language models
• Mixture-of-Transformer-Experts (MoT) architecture

Use Cases

• Describing and understanding images (e.g., 'Tell me about this picture')
• Generating photorealistic images from text prompts (e.g., 'a photo of three antique glass magic potions')
• Editing images while preserving details (e.g., 'He squatted down and touched a dog's head')
• Transforming image styles (e.g., 'Change to 3D animated style')
• Navigating and interacting with virtual environments (e.g., 'After 0.40s, move forward')
• Engaging in multi-turn conversations with compositional reasoning (e.g., creating a slogan for a doll)
• Refining prompts for detailed and coherent visual outputs using a 'thinking' mode