From Prompt to Polygon: Evaluating End-to-End Human–AI Workflows for Web-Based Generative 3D Authoring

Prompt- and image-driven 3D generation can lower the interaction barrier to digital asset creation, yet model-level advances do not establish whether non-specialists can operate a complete authoring workflow. This study designed and evaluated a web-based human–AI prototype integrating a React interface, FastAPI services, ComfyUI workflow orchestration, asynchronous job processing, persistent asset storage, and browser-based 3D visualization. We implemented four modes: Text-to-3D, Text-to-3D with ControlNet, Single Image-to-3D, and Multi Image-to-3D. Evaluation combined 14 requirement-derived black-box scenarios, cross-viewer GLB compatibility checks in three viewing environments, 35 completed generation jobs, task-based observation with seven participants completing 84 task executions, System Usability Scale (SUS) responses from a separate group of 12 students, and one expert assessment of implementation and output quality. All 14 functional scenarios produced the expected outcomes, and the tested GLB outputs opened in the integrated viewer, Blender, and 3dviewer.net. Mean server-side turnaround was 66.15 s; an exploratory Kruskal–Wallis test indicated differences among the observed workload distributions (H(3) = 10.52, p = .015, ε² = .243), but did not support causal or pairwise speed claims. All observed tasks were eventually completed, but only 61 of 84 executions (72.62%) were independent; Tasks 5 (enter Studio) and 10 (find the generated model in My Projects) each required assistance for five of seven participants. Mean SUS was 81.67 (SD = 9.25; 95% CI [75.79, 87.55]), while generated-model quality received a diagnostic expert score of 3.50/5. The study contributes an end-to-end human–AI authoring architecture and a multidimensional evaluation logic that separates functional conformance, deployment turnaround, cross-viewer compatibility, observed interaction, perceived usability, and output quality. Findings support the feasibility of accessible web-based generative 3D authoring while identifying discoverability and geometric-quality constraints.