HyperTransport: Amortized Conditioning for T2I Models
A new method called HyperTransport uses a hypernetwork to amortize the cost of activation steering in text-to-image generative models. Existing steering techniques require per-concept optimization, making them slow for large or evolving concept sets. HyperTransport maps embeddings from a pretrained CLIP encoder directly to intervention parameters, trained end-to-end, enabling fast conditioning at request time without fine-tuning.
Key facts
- HyperTransport is a hypernetwork framework for amortized conditioning.
- It addresses the brittleness of prompting and high cost of fine-tuning.
- Existing activation steering methods need per-concept optimization.
- HyperTransport maps CLIP embeddings to intervention parameters.
- The method is trained end-to-end.
- It enables fast deployment for large or evolving concept sets.
- The paper is on arXiv with ID 2605.08254.
- The approach targets text-to-image generative models.
Entities
Institutions
- arXiv
- CLIP