Tiny-Engram: Trigger-Indexed Concept Tables for Generative Vision
Tiny-Engram has been unveiled by researchers as a compact, trigger-indexed concept table that assigns explicit lexical addresses and activation boundaries to visual memories within frozen image and video generators. Each concept is represented by a small collection of memory entries that are indexed through registered n-gram matches, influencing text-encoder hidden states solely within the designated trigger area. The conditioning pathway remains unchanged outside this lexical framework, adhering to the frozen base model. This innovative approach links a specific trigger phrase to a target identity while maintaining compositional control from the surrounding prompt. Tiny-Engram addresses a significant shortcoming of existing personalization techniques for generative vision models, allowing for explicit management of concept retrieval in personalized content generation.
Key facts
- Tiny-Engram is a compact trigger-indexed concept table for generative vision models.
- It gives visual memories explicit lexical addresses and activation boundaries.
- The method works with frozen image and video generators.
- Each concept is parameterized as a small set of memory entries indexed by n-gram matches.
- Modulation occurs only within the matched trigger region.
- Outside the trigger region, the conditioning pathway is identical to the frozen base model.
- The method works with both single-encoder latent diffusion and multi-encoder diffusion-transformer backbones.
- It binds a rare trigger phrase to a target identity while preserving compositional control.
Entities
Institutions
- arXiv