per-subject training runs were the 2023-2025 route to identity control; 2026 production models (Seedance 2.5, MiniMax H3, Seedream 4.5) hold identity from reference images supplied in-context at gener
Learn more →
Why study this historical topic?
per-subject training runs were the 2023-2025 route to identity control; 2026 production models (Seedance 2.5, MiniMax H3, Seedream 4.5) hold identity from reference images supplied in-context at gener
Subject Personalization of Text-to-Image Models (Textual Inversion and DreamBooth)
Techniques that teach a pretrained text-to-image model a new specific subject from a few reference images so it can be rendered in novel scenes: per-subject optimization of a new word embedding (Textual Inversion) or model weights (DreamBooth), and faster encoder-based approaches that embed the subject in a forward pass, all trading off identity fidelity against preserving the model's prior.
DreamBooth: Finetuning Large Diffusion models! Research Paper Explained
DreamBooth is a fine-tuning technique for large diffusion models designed to inject subject-specific representations into a pre-trained model while preserving general class priors via regularization.…