DreamReader: An Interpretability Toolkit for Text-to-Image Models
Nirmalendu Prakash, Narmeen Oozeer, Michael Lan, Luka Samkharadze, Phillip Howard, Roy Ka-Wei Lee, Dhruv Nathawani, Shivam Raval, Amirali Abdullah · 2026
Read DreamReader: An Interpretability Toolkit for Text-to-Image Models by Nirmalendu Prakash, Narmeen Oozeer, Michael Lan, Luka Samkharadze, Phillip Howard, Roy Ka-Wei Lee, Dhruv Nathawani, Shivam Raval, Amirali Abdullah (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13299.txt.
Description
Nirmalendu Prakash and co-authors release DreamReader, a toolkit for studying text-to-image diffusion models. Work on how these models work inside has been split into lone tests with no shared frame. DreamReader joins activation extraction, causal patching, ablations, and steering in one system. It works across model types and time steps, with tests run in set batches. Three new tools let users tune parts of a model and steer it with probe scores. Case studies show LLM-born methods can steer image models in a controlled way.
Catalog record
| CREATOR | Nirmalendu Prakash, Narmeen Oozeer, Michael Lan, Luka Samkharadze, Phillip Howard, Roy Ka-Wei Lee, Dhruv Nathawani, Shivam Raval, Amirali Abdullah |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28411 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.13299