ReContraster Beats Every Poster AI by Teaching Contrast Like a Designer
A training-free system uses three cooperating LLM agents plus a two-stage diffusion trick to generate posters with strong visual contrast between regions.
ProTakeaways
- ReContraster is a training-free poster generator that engineers strong regional contrast in the output.
- Three LLM agents (cognition, arranger, refiner) iterate on element pairs, layout, and quality checks.
- A hybrid denoising strategy uses gradient consistency loss plus joint region denoising to hide seams.
- Beats SD3, Flux.1-dev, OpenCOLE, PosterGen, and PosterMaker on all seven metrics evaluated.
- Ships with a new 643-poster benchmark annotated using GPT-4o and SAM, to be released on publication.
- Accepted to ACL 2026; no code available yet.
Most text-to-poster systems can nail a clean layout, but the results still look flat. A new paper from Beijing University of Posts and Telecommunications and Peking University argues that the missing ingredient is contrast, the visual tension between a burning forest and a green one, or a glacier and a desert, and proposes a training-free pipeline called ReContraster that engineers that tension explicitly.
ReContraster is described as the first training-free model to leverage regional contrast for poster generation, emulating a designer's workflow through a compositional multi-agent system that identifies elements, organizes layout, and evaluates candidates. The paper was accepted to ACL 2026.

Why pretty posters still fall flat
Poster design has a job most images do not: it has to grab someone in a fraction of a second. The authors lean on the psychological principle of "contrast effects," which says the brain fires harder when it sees stark differences between adjacent stimuli. Existing systems like SD3, Flux.1-dev, OpenCOLE, PosterGen and PosterMaker can produce coherent layouts and even readable text, but they tend to either generate two regions that look nearly identical in color and content, or slam contrasting scenes together with an ugly seam down the middle.
ReContraster targets three properties at once: element contrast (choosing opposing objects and colors), aesthetic harmony (keeping the composition balanced), and boundary coherence (making transitions between regions look intentional rather than pasted).