Collaposer: Transforming Photo Collections into Visual Assets for Storytelling with Collages

TL;DR

Collaposer tackles the laborious prep work in collage storytelling by turning photo collections into organized, story-aligned visual assets. It combines open-set tagging, instance segmentation, and LLM-driven semantic association to extract and arrange object-level elements into a hierarchical, actionable asset library; assets are scored and presented by content diversity, story consistency, and resolution to support nonlinear, rapid collage creation. A formative study informs design goals, and a user study demonstrates that Collaposer outperforms ablated baselines in asset-story consistency, diversity, and usability, while enabling one-pass static collage creation and easy export for animation pipelines. The tool promises to shift creator effort from manual search and segmentation toward composition and storytelling, with practical implications for rapid ideation and scalable narrative asset management in digital collage and related visual storytelling workflows.

Abstract

Digital collage is an artistic practice that combines image cutouts to tell stories. However, preparing cutouts from a set of photos remains a tedious and time-consuming task. A formative study identified three main challenges: 1) inefficient search for relevant photos, 2) manual image cutout, and 3) difficulty in organizing large sets of cutouts. To meet these challenges and facilitate asset preparation for collage, we propose Collaposer, a tool that transforms a collection of photos into organized, ready-to-use visual cutouts based on user-provided story descriptions. Collaposer tags, detects, and segments photos, and then uses an LLM to select central and related labels based on the user-provided story description. Collaposer presents the resulting visuals in varying sizes, clustered according to semantic hierarchy. Our evaluation shows that Collaposer effectively automates the preparation process to produce diverse sets of visual cutouts adhering to the storyline, allowing users to focus on collaging these assets for storytelling. Project website: https://jiayzhou.github.io/collaposer-website/

Figures (6)

• Figure 1: A comparison between manual workflows and our approach for asset preparation in collage-based storytelling. Our formative study identified three challenges in manual workflows: inefficient keyword-based photo search, tedious manual image cutout, and difficulty in organizing large collections of cutouts through manual grouping and naming. Collaposer introduces three corresponding AI-driven features that alleviate the challenges: inference of central and related elements based on the story, automatic instance segmentation, and asset management with semantic hierarchy.
• Figure 2: Our pipeline consists of three stages. The inputs include an image collection and a story description. In Stage I, valid visual elements are trimmed out and tagged with an object name. In Stage II, visual elements relevant to the story are selected and clustered into semantic groups. The elements classified as characters undergo part segmentation and pose estimation for later manipulation. In Stage III, the visual assets are presented in a compact view to facilitate navigation and composition.
• Figure 3: A user's workflow with Collaposer. In the source panel, (a) the user selects a photo collection and (b) browses the collection. Then, the user (c) enters a story description in the input box and clicks the submit button. The system interface transitions to the layer panel when the preparation of visual assets is completed. (d) Collaposer presents the visual assets in the layer panel. (e) User interacts with the tree view and the canvas view to pick up and arrange visual elements. After they are satisfied with the static collage story, (f) the user clicks the export button to obtain two JSON files. (g) They run a script and edit the collage story as clustered layers.
• Figure 4: User ratings across four evaluation dimensions—Consistency, Diversity, Presentation, and Usability—covering eleven question items (Q1–Q4) for three system variants: Collaposer (c), Ablated-Select (as), and Ablated-Present (ap). Ratings were collected on a 7-point Likert scale (1 = Strongly Disagree, 7 = Strongly Agree). Asterisks (*) indicate statistically significant differences in mean ratings (p < .05 / 3). Overall, the results indicate that Collaposer supports more effective story-aligned asset selection and presentation compared to the baselines. Ablated-Select shows the weakest asset-story consistency, often failing to provide relevant elements and occasionally including unrelated ones. Ablated-Present delivers the least satisfactory presentation results, though it has a relatively smaller impact on system usability.
• Figure 5: Collaposer's selection aligns with P5's mental visualization of the story and enriches some details. P12's story created with Collaposer starts from a broad description "different animals are playing on the playground" and adds details through the composition of visual elements.