bytedance-seed/bagel

BAGEL is an open-source AI model from ByteDance's research team that can both understand and generate images alongside text, using a single unified model. Most AI image tools either analyze images or create them, but BAGEL does both within one system, plus more advanced tasks like editing existing photos, generating multi-angle 3D views from a single image, and predicting how a scene might look after actions are taken. The model has 7 billion active parameters (14 billion total) and was trained on a large mix of text, image, video, and web content. In standard tests comparing AI image models, BAGEL scores competitively against other leading open-source models for understanding images (such as answering questions about photo content), while also producing image generation quality that stands alongside dedicated image-generation tools. For people who want to run it themselves, the project provides Python scripts covering several tasks: generating an image from a text description, editing an existing image based on instructions (such as removing the background or changing the sky), and chatting about what is in an image. The model requires a capable GPU with a large amount of memory (around 80 GB at full precision). Community members have released compressed versions that use less memory, and the project includes Docker setup files and a Windows installation guide. Researchers and developers can also access a live demo site and a Hugging Face Space to try the model without installing anything. The training process, benchmark evaluation code, and model weights are all publicly available. Recent updates include new evaluation benchmarks, community-contributed compression tools, and a ComfyUI plugin for no-code image workflows. If you want to inspect or modify the model itself, the architecture is described in a published research paper linked from the README. The project also includes a Discord community for troubleshooting and sharing results.