From Prompting to Composing: A Spatial Canvas Interface for Poster Generation

From Prompting to Composing. A Spatial Canvas Interface for Poster Generation. By Yitong Wang and seven others, from Fudan University, Microsoft Research, and four other institutions. Published on the eighth of October.

A poster is a two-dimensional thing. Where the title sits, where the product goes, which corner holds the logo: that is composition, and it lives in space. Text prompting asks users to encode that inherently two-dimensional intent into a one-dimensional sequence of words. The authors call prompting an indirect interface for poster generation, and the indirection is the problem. Layout has to survive a trip through a sentence before the model ever sees it.

The proposal is to let users compose intent directly in space. A Spatial Canvas Interface does that through four complementary binding types: semantic, identity, text, and pixel. Alongside them sit Text Specifications, which describe individual elements and the global appearance of the poster. The canvas says where, and the specifications say what, so intent is specified separately from the act of generating pixels.

On top of this interface the authors build Compo, a poster generation model adapted from a pretrained image editing model so that it understands Spatial Canvas inputs and Text Specifications. It runs in two modes. In direct inference, the user constructs the canvas explicitly. In agentic mode, a high-level request is automatically translated into a planned Spatial Canvas, so a user can start from a sentence and still end up with a composed layout.

Training Compo relies on a scalable pipeline that automatically constructs supervision data for each binding type and for their combinations. That is what allows efficient adaptation of an existing model, without training a specialised poster generator from scratch. To measure the result, the authors also introduce a benchmark that checks adherence to each binding type on its own and to their joint composition.

Experiments report that Compo achieves stronger compositional controllability than two kinds of competitor: general-purpose image generation models, and dedicated poster generation systems. That gain in control comes while maintaining high visual quality, so control is not bought by giving up how the poster looks.

The title carries the argument. Prompting treats a poster as something you describe and hope for. Composing treats it as something you arrange, with the generator responsible only for rendering what was arranged. Decoupling intent specification from visual generation is the move, and the agentic mode shows it does not have to cost the convenience of a one-line request.