Identity tells the system who. Style tells the system how to see. Place tells the system what the world is.
PRISM is a closed-loop, place- and retrieval-informed scene memory system for image generation. It separates generation into three independent, retrievable memory systems -- identity, style, and place -- and verifies its own output against place memory before delivering a result.
- White Paper -- Full technical architecture, implementation status, limitations, and evaluation framework
- Pitch -- 6-page overview of the problem, approach, and differentiators
- Three-memory architecture (identity, style, place) with structured retrieval
- Closed-loop geographic verification: generate, verify, strengthen, re-generate
- Zero-photo autonomous place construction from a location name
- Hierarchical place memory with geographic fallback
- GeoGen-Bench evaluation framework (25 US locations, 4 methods, 3 evaluators)
- VLM-powered automated analysis pipeline for building memory at scale
- Model-agnostic: supports API-based VLMs (Claude) and locally-hosted models via vLLM
The architecture is built. The benchmark framework is ready. Quantitative evaluation results are next. See the Implementation Status section in the white paper for a detailed component-by-component breakdown.
Copyright © 2026 Anthony O'Dwyer. Licensed under CC BY-NC-ND 4.0.