PosterMate: AI-Assisted Poster Design
- PosterMate is an AI-assisted poster design system that leverages simulated target audience personas to generate iterative and diverse feedback.
- It integrates marketing briefs with moderator-led discussions to reconcile differing perspectives and adapt poster elements effectively.
- The system employs multimodal strategies and JSON-level editing to refine text, image, and theme components while preserving designer agency.
Searching arXiv for PosterMate and closely related poster-generation/editing systems to ground the article with current papers. PosterMate is an AI-assisted poster design system for advertisement poster design that facilitates collaboration by creating audience-driven persona agents constructed from marketing documents, gathering feedback from each persona agent regarding poster components, and stimulating discussion with the help of a moderator to reach a conclusion; the agreed-upon edits can then be directly integrated into the poster design (Shin et al., 24 Jul 2025). It is motivated by the observation that poster design benefits from synchronous feedback from target audiences, while recruiting relevant audiences, obtaining iterative critique, and reconciling heterogeneous perspectives are all difficult in practice. Within that framing, PosterMate is not principally a poster generator; it is a system for audience-aware critique, negotiation, and editable revision.
1. Design problem and conceptual scope
PosterMate addresses a collaboration problem rather than a pure generation problem. In advertisement poster design, the success of a poster depends on alignment with intended viewers, yet synchronous access to diverse target audiences is expensive and slow. The paper therefore treats recent generative AI models as a means to simulate human-like interactions in feedback processes, and specifically uses persona agents to represent different target audience segments derived from a marketing brief (Shin et al., 24 Jul 2025).
The system is explicitly grounded in campaign materials already present in commercial workflows. A user uploads a marketing brief and works from an initial poster draft on a canvas; PosterMate extracts campaign context, generates persona agents, solicits their critiques of the draft, and supports either direct adoption of individual suggestions or moderator-led reconciliation of conflicting recommendations. This makes the system closer to a collaborative design assistant than to a one-shot creative generator.
A recurrent misconception is that PosterMate automates poster production end to end. The paper states that it is not a fully automatic poster generator and is better described as a semi-automatic, human-in-the-loop editing system. The user chooses whether to apply a persona’s feedback, apply the moderator’s conclusion, ignore it, refine it through more discussion, or manually edit the canvas. This design preserves designer agency while embedding audience simulation directly into the editing environment.
2. Persona synthesis from marketing briefs
PosterMate’s persona construction pipeline begins from a marketing brief, because such briefs already encode campaign goals, target-audience descriptions, and practical constraints. The paper argues that briefs usually provide only high-level audience summaries, so the system refines them into diverse audience segments through a two-dimensional persona space (Shin et al., 24 Jul 2025).
The model identifies two steerable dimensions whose extremes do not contradict the target audience descriptions already present in the brief, then instantiates the four combinations of those extremes. This yields four persona agents. The choice of a 2×2 structure is deliberate: it is common in business settings, provides diversity without excessive complexity, and avoids the combinatorial growth that would follow from adding more dimensions.
Each persona is represented through a fixed qualitative schema consisting of: name, one-line summary, background, goal/motivation, challenge/pain point, need, one-line quote, and one-line rationale explaining how the persona’s perspective could contribute to poster design. The system also extracts a high-level marketing goal from the brief and generates an avatar from the persona’s name to increase social affordance and co-presence.
This design implies that PosterMate’s “audience model” is not a latent embedding alone but a structured textual profile used repeatedly in downstream prompting. A plausible implication is that the quality of critique depends not only on the LLM backbone but also on whether the brief contains sufficiently rich and non-contradictory audience cues.
3. Feedback generation, conflict detection, and moderated discussion
PosterMate organizes critique around three poster component categories: text, image, and theme (Shin et al., 24 Jul 2025). This division was motivated by prior poster-design literature and by the system’s formative study, which found that users preferred holistic theme feedback over more fragmented style-only critique.
Each persona agent receives the persona description, the poster design goal, the raw marketing brief, and the poster in two synchronized forms: a rendered pixel image and serialized JSON exported from Polotno. The rendered image supports appearance-level reasoning, while the JSON supports component-level targeting and later integration.
For text components, the output contains a target component identifier, a high-level opinion, and a preview containing the revised text. For image components, the output contains a target image component identifier, a high-level opinion, and a preview in the form of a one-line image description. For theme feedback, the output contains a one-line opinion of the desired mood plus two theme descriptors, tone and color. These descriptors are then used in theme recommendation.
Theme recommendation uses Polotno’s template library with over 1,000 public templates. Each template is embedded using ViT-base-patch16-224-in21k and stored in vector storage. At runtime, PosterMate takes the tone and color descriptors, uses a text-to-image model to generate a poster-like image reflecting them, embeds that image with the same ViT model, computes cosine similarity against all template embeddings, and ranks templates by similarity (Shin et al., 24 Jul 2025).
When conflicting recommendations arise, PosterMate groups feedback by component identifier, asks an LLM to detect conflicting feedback with distinct focuses, and generates a one-line summary of the conflict. A moderator agent then asks the user whether they have any opinion or preference, generates a thought-provoking question for each persona agent, collects persona responses in parallel, and produces a conclusion consisting of the component identifier, a summary of the conclusion, and a preview. The paper states that persona agents are instructed to remain open-minded and collaborate toward compromise. If disagreement remains, the moderator may selectively omit certain portions of some agents’ perspectives in order to arrive at a conclusion that satisfies the majority.
| Component | Persona output | Accepted integration |
|---|---|---|
| Text | Opinion + revised text preview | Update target text property |
| Image | Opinion + one-line image description | Generate image and replace src |
| Theme | Mood opinion + tone and color descriptors | Retrieve and adapt a template |
This moderator-centered mechanism is the system’s central social abstraction. Rather than reducing multiple critiques to a score or a vote, it turns conflict into explicit deliberation.
4. Editable poster integration and system architecture
PosterMate integrates accepted edits directly into the poster canvas through JSON-level manipulation (Shin et al., 24 Jul 2025). If text feedback is accepted, the system locates the component by identifier and updates its text property with the preview content. If image feedback is accepted, it identifies the target image component, generates an image from the one-line description using a text-to-image model, and replaces the component’s src property with the generated image URL.
Theme transfer is more involved. The user selects a recommended template from the ranked list, after which the system takes the JSON of the current poster and the JSON of the selected template and prompts a model to map original text and image components onto corresponding template elements. If the template has extra or missing components, it adds or removes them accordingly. Theme embellishments such as SVGs are temporarily stored separately with their z-index values so that visual hierarchy can be preserved during reassembly.
After theme adaptation, the system checks whether layout fixes are needed. Using both the rendered poster and the JSON structure, the model detects overlap and suggests repositioning or resizing by returning affected component identifiers and one-line descriptions of the required changes. The JSON is then updated and re-rendered. This means theme application is not merely template substitution; it is a constrained remapping followed by layout repair.
The implementation stack is explicit. PosterMate uses OpenAI GPT-4o as the multimodal LLM, DALL-E 3 for text-to-image generation, ViT-base-patch16-224-in21k for template embedding, Polotno as the canvas/editor and JSON import/export layer, SvelteKit for the frontend, and FastAPI for the backend (Shin et al., 24 Jul 2025). The use of both image and JSON representations is a notable systems decision: it supports visual reasoning, exact component manipulation, and post-edit overlap checking within a single workflow.
5. Empirical evaluation and findings
PosterMate was evaluated through both an end-to-end user study and a controlled online evaluation (Shin et al., 24 Jul 2025). The user study had participants. Through that study, the paper reports that PosterMate captured overlooked viewpoints and served as an effective prototyping tool. Participants used the discussion feature actively, with an average of 4.3 discussions and , and qualitative analysis indicated that the system helped surface audience perspectives that designers would otherwise miss.
The controlled online evaluation had participants recruited on Prolific. It examined two questions: whether the moderator’s conclusion effectively synthesized the viewpoints of multiple personas, and whether the feedback from an individual persona agent was appropriate given its persona identity.
For synthesis quality, evaluators saw outputs from four individual personas and one moderator conclusion for each data type. The moderator conclusion was selected as most likely to satisfy the majority of personas in 40.5% of text cases (), 69.5% of image cases (), and 42.5% of theme cases (). The corresponding averages for individual persona feedback were 14.9% for text (), 7.6% for image (), and 14.3% for theme (). The reported chi-square and post-hoc tests were significant across tasks, supporting the claim that discussion effectively synthesizes different persona agents’ perspectives (Shin et al., 24 Jul 2025).
For persona-identity appropriateness, evaluators were asked to match outputs to personas. Correct matching rates were 52.1% for text (), 64.3% for image (0), and 21.9% for theme (1), against a 25.0% chance rate. The paper interprets this as evidence that text and image feedback were recognizably persona-appropriate, while theme feedback was less clearly attributable to specific personas.
The evaluation also identified limitations. Theme alignment was the weakest part of the system; evaluators had difficulty linking themes to personas, and the authors suggest that tone/color descriptors did not always map cleanly to retrieved templates. The controlled study also omitted user comments for consistency, only two marketing briefs were used across the studies, and the paper does not directly compare final posters produced with PosterMate against posters improved without AI or with alternative AI tools. These caveats matter because they constrain claims about generalizability and causal attribution.
6. Position within AI-assisted poster design
PosterMate occupies a distinct position within the broader poster-design literature. Systems such as “P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark” generate HTML-rendered academic posters directly from research papers (Sun et al., 21 May 2025), while “APEX: Academic Poster Editing Agentic Expert” formalizes interactive academic poster editing from a source poster, source paper, and user instruction (Shi et al., 8 Jan 2026). “CreatiPoster: Towards Editable and Controllable Multi-Layer Graphic Design Generation” instead generates editable, multi-layer compositions from optional natural-language instructions or assets (Zhang et al., 12 Jun 2025). PosterMate, by contrast, is centered on audience simulation, critique, and moderated reconciliation rather than on automatic poster generation or general-purpose graphical editing (Shin et al., 24 Jul 2025).
This suggests that PosterMate’s novelty lies in reframing AI support around simulated stakeholder ecosystems inside the editor. Rather than asking a model to directly optimize a poster, it operationalizes audience plurality: multiple personas critique the draft, disagreements are surfaced explicitly, and a moderator produces a compromise-oriented conclusion that remains editable and optional. In that sense, the system is less a generator than a socio-technical coordination layer for design feedback.
The paper also identifies broader risks and open questions. Persona generation from broad audience descriptions may reinforce stereotypes or hallucinated preferences, and users may over-trust AI feedback because it appears reasoned and socially embodied. At the same time, the system demonstrates that AI can support collaborative critique rather than merely automated creation. For design research, that makes PosterMate a notable example of human-AI co-design organized around feedback, explanation, and negotiation instead of one-shot generation.