Image Generation from Contextually-Contradictory Prompts DOI: XXXXXXX.XXXXXXX 1201
Saar Huberman, Or Patashnik, Omer Dahary, Ron Mokady, Daniel Cohen-Or · 2025
Read Image Generation from Contextually-Contradictory Prompts DOI: XXXXXXX.XXXXXXX 1201 by Saar Huberman, Or Patashnik, Omer Dahary, Ron Mokady, Daniel Cohen-Or (2025) 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-2506-01929.txt.
Description
Saar Huberman and co-authors tackle, in a 2025 paper, a failure mode of text to image models: prompts whose concepts quietly contradict each other. A diffusion model may have learned links so tight that one idea in the prompt cancels another. The result looks fine as an image, but misses what was asked.
Their fix works stage by stage. Image building in these models proceeds from noise to detail, and different content emerges at each stage. The team builds a chain of proxy prompts, each matched to what should appear at its stage, and each kept free of the clash. A large language model reads the target prompt, spots the contradiction, and writes the proxies in plainer terms that keep the intent. In tests across hard prompts, the method gave more faithful images. The approach needs no retraining of the image model.
Catalog record
| CREATOR | Saar Huberman, Or Patashnik, Omer Dahary, Ron Mokady, Daniel Cohen-Or |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-01929