The study followed a software design-and-evaluation approach based on the Prototype Model of the software development life cycle [20]. Iterative prototyping was appropriate because the application joins computationally intensive generation, asynchronous job control, sequential workflow stages, and user-facing feedback. Development followed communication, quick planning, rapid design, prototype construction, and deployment-feedback cycles. The research evaluated the integrated application rather than training or benchmarking a new generative model.

The unit of evaluation differed across the technical measures. Functional evidence came from requirement-derived scenarios, compatibility evidence from opening exported files in independent viewing environments, and performance evidence from completed jobs. Human-centered evidence came from direct task observation, SUS responses, and one expert assessment. Separating these forms of evidence prevents a functional pass rate or favorable SUS score from being interpreted as proof of geometric fidelity, production reliability, or learning impact.

The prototype uses a client–server architecture with asynchronous processing. The React frontend collects text or image inputs, displays job state, and renders generated 3D assets, while the FastAPI backend validates requests, selects a workflow, injects parameters, and coordinates result delivery. ComfyUI executes the selected generative graph, PostgreSQL stores application and job metadata, and ImageKit stores generated assets. Figure 1 maps the user-facing path from input and job submission to workflow execution, persisted output, and browser-based inspection.

Evaluation drew on six evidence sources. Fourteen black-box scenarios examined expected input–output behavior, and compatibility checks opened exported GLB files in the integrated viewer, Blender, and 3dviewer.net. Performance analysis used 35 completed jobs: 20 Text-to-3D jobs and five jobs for each of the ControlNet, Single Image-to-3D, and Multi Image-to-3D modes. Task-based testing involved seven convenience-sampled students, five from Visual Communication Design, one from Civil Engineering, and one from Informatics, who each completed 12 tasks, yielding 84 task executions; SUS was completed by 12 respondents.

The implementation used React with Vite for the frontend; Python and FastAPI for backend services; Clerk for authentication; ComfyUI, PyTorch, and Diffusers for generative execution; Stable Diffusion for text-to-image synthesis; foreground removal before reconstruction; Hunyuan3D for image-to-3D generation; PostgreSQL with SQLAlchemy for application data; ImageKit for asset storage; and React Three Fiber, Drei, and Three.js for GLB visualization. ControlNet supplied an additional spatial-conditioning mechanism in the corresponding mode [21]. The broader technical context includes implicit occupancy representations [22], continuous signed-distance functions [23], differentiable volumetric rendering [24], sparse-input radiance fields [25], neural mesh rendering [26], and multiresolution hash encodings [27]. These works motivate component choices but are not treated as models newly proposed or retrained in this study.

The retained research record identifies the software families and deployed functions but does not preserve exact package versions, GPU model, CUDA version, random seeds, or dependency lock files. These omissions limit environment-level reproducibility and prevent hardware-normalized performance comparison. Turnaround is therefore reported only for the observed deployment; the article does not claim portable inference speed or exact computational reproducibility.

The expert validator was selected for subject-matter expertise in artificial intelligence, machine learning, and 3D animation. System implementation was assessed with 20 five-point Likert items (1 = strongly disagree to 5 = strongly agree) covering function and flow, input and validation, job management and status, AI workflow integration, multi-stage processing, prompt enhancement, output, visualization, viewer interaction, download, history, gallery, authentication, user-data separation, credits, interface clarity, and overall feasibility. Generated-model quality was assessed separately with 10 five-point items covering geometry, structural completeness, shape accuracy, detail, surface quality, multiview consistency, visual conformity, prompt/image correspondence, 3D-asset suitability, and GLB-output suitability. Appendix Table D1 reports system-implementation scores, and Appendix Table D2 reports generated-model quality scores. One validator completed both instruments; consequently, inter-rater reliability cannot be estimated and the results are diagnostic rather than a validated population-level quality estimate.

The main Text-to-3D path converts a prompt into a 2D image, removes its background, reconstructs a 3D mesh, and exports the result as GLB. The backend selects a reusable workflow template according to the requested mode and injects prompt, negative prompt, seed, sampling steps, CFG scale, and image inputs where applicable. Each request is represented as an asynchronous job. After the workflow is sent to ComfyUI, the returned prompt_id links the ComfyUI execution to the application job record; the browser polls job state until completion or failure. On completion, the backend retrieves the GLB and thumbnail, uploads them to ImageKit, records persistent asset URLs in PostgreSQL, and makes the result available for interactive viewing and download.

Single Image-to-3D and Multi Image-to-3D accept user-provided images rather than a synthesized image. Multiple views can add geometric evidence: 3D-R2N2 addresses single- and multi-view reconstruction [14], while differentiable ray consistency uses multi-view observations to supervise single-view prediction [28]. Single-image methods such as Zero-1-to-3 [11], pixelNeRF [12], and PIFu [13] must infer unseen geometry. The four modes are therefore distinct input and conditioning paths that converge on the same job-management and GLB-delivery mechanism.

A functional scenario passed when observed behavior matched the requirement-derived expected result. The complete 14-scenario matrix is provided in Appendix Table A1, including a negative empty-prompt validation case and a simulated process-failure case. The retained record documents these tests as part of the integrated prototype evaluation but does not preserve an individual tester identity or an immutable build hash; accordingly, the paper does not claim a formally frozen release build.

Cross-viewer GLB compatibility was defined narrowly as the ability of the tested GLB output to open and display without a reported read/import failure in the integrated website viewer, Blender, and 3dviewer.net. This is not treated as full interoperability because material preservation, texture integrity, transforms, mesh validity, schema conformance, and repeated import–export behavior were not evaluated.

For performance, processing_time was defined as server-side turnaround rather than isolated model inference. It includes workflow monitoring, artifact retrieval from ComfyUI, and transfer to ImageKit. Mean, median, minimum, and maximum were reported by mode, and only jobs recorded as COMPLETED were available for analysis. One retained job required an internal retry before completion; an overall generation success rate is not calculated because the archive does not contain a complete denominator of all attempts, final failures, and cancellations.

Task observation used the 12-task script and completion records presented in Appendix Table B1. Each task execution was recorded as independent completion, completion with assistance, or failure. The retained task-level record provides independent and assisted completion counts and error events by task, as well as participant-level session times (Appendix Table B2); it does not preserve separate time-on-task values for each task or a sufficiently detailed intervention threshold to reproduce the facilitator’s assistance decision exactly. Observed errors were categorized as input, navigation, interpretation, or procedural. Because the task-based and SUS groups were explicitly separate, no participant-level overlap is assumed.

The seven-person task study is treated as formative diagnostic usability testing rather than a summative population estimate. Early usability research showed that small formative samples can reveal recurrent interface problems with diminishing returns [29], while later evidence demonstrates substantial variability in what very small samples detect and cautions against treating a five-user rule as universally sufficient [30]. The present sample is therefore used to locate observed friction, not to estimate its prevalence in a broader population.

SUS used the standard 10-item scoring procedure [31]: odd-item responses were reduced by one, even-item responses were subtracted from five, the adjusted items were summed, and the result was multiplied by 2.5. The original study protocol classified scores of at least 70 as Acceptable; this label is used descriptively because SUS interpretation is context-dependent [32].

Because the performance groups were small and unbalanced, an exploratory Kruskal–Wallis test compared turnaround distributions across the four modes at α = .05. Epsilon squared (ε²) summarized omnibus effect magnitude, and no pairwise post hoc claim was made. Mode, input complexity, system load, workflow composition, and transfer overhead are not separable in the retained deployment data; therefore, statistical significance is not interpreted as evidence that generation mode caused the observed timing differences.

SUS results were summarized by the mean, sample standard deviation, median, range, and a two-sided 95% t confidence interval for the mean. Task executions were not treated as independent participants, so no confidence interval was attached to the 84 execution-level outcomes. Functional scenarios support conformance only for executed cases; compatibility supports file opening rather than geometric correctness; completed-job timing does not estimate failure probability; and the single expert assessment is an initial judgment rather than an inter-rater reliability study. No claim of production scalability, educational effectiveness, or state-of-the-art reconstruction follows from these measures.

All 14 requirement-derived black-box scenarios produced the expected outcomes (Appendix Table A1). The matrix covers prompt submission, empty-input validation, job creation, queued–processing–completed transitions, failure messaging, generation completion, viewer loading and interaction, download, Text-to-3D multi-stage execution, Single Image-to-3D, Multi Image-to-3D, ControlNet, and periodic status polling. These results demonstrate functional conformance for the executed cases, not exhaustive defect absence.

The tested GLB output opened successfully in the website viewer, Blender, and 3dviewer.net. This supports cross-viewer GLB compatibility, but not full interoperability because material preservation, mesh validity, transforms, schema conformance, and round-trip exchange were not evaluated.

The performance dataset contained 35 jobs recorded as COMPLETED: 20 Text-to-3D jobs and five jobs in each remaining mode. Table 1 summarizes server-side turnaround by mode. Because prompts and images were heterogeneous and mode counts were unequal, the statistics characterize the observed workload rather than a controlled model benchmark.

Across all jobs, mean turnaround was 66.15 s and the median was 63.65 s. The exploratory omnibus test indicated that at least one observed workload distribution differed, H(3) = 10.52, p = .015, ε² = .243. This result does not identify a superior mode or establish mode as the cause of timing differences. Input complexity, workflow composition, shared infrastructure, transfer overhead, and unrecorded system load may contribute to the observed variation.

Single Image-to-3D included one 10.87 s observation but otherwise clustered above 53 s, whereas ControlNet and Multi Image-to-3D extended beyond 108 s. These values are end-to-end observations from one deployment rather than hardware-normalized inference latency. The asynchronous architecture is nevertheless supported pragmatically by the observed tens-of-seconds turnaround because it allows the browser to return promptly and communicate status while the backend continues processing.

The examples retained recognizable characteristics of their prompts or input images, but geometric detail, structural completeness, and input correspondence varied. This variation is consistent with the expert assessment and cautions against equating successful file generation with consistently high-quality geometry. No reference mesh, multiview ground truth, or blinded perceptual comparison was available. The visual evidence therefore supports feasibility and identifies failure dimensions without quantifying reconstruction accuracy.

Seven participants each attempted 12 tasks, producing 84 task executions. Every task was eventually completed, but 61 executions (72.62%) were independent, and 23 (27.38%) required assistance. No execution was classified as a failure. Table 2 shows task-level independent and assisted completion, showing that assistance was concentrated in particular tasks.

Tasks 5 (enter Studio) and 10 (find the generated model in My Projects) each required assistance for five of seven participants, while Tasks 9 (make the model public) and 12 (log out) each required assistance for three participants. This task-level pattern makes the global 27.38% assistance rate actionable: the main issue was not inability to complete the workflow, but discoverability of navigation and result-management actions.

Forty-eight observed error events were recorded across the 84 executions: 28 input errors (58.33%), 16 navigation errors (33.33%), two interpretation errors (4.17%), and two procedural errors (4.17%). Error events were concentrated in Task 1 (12 events), Task 10 (9), Tasks 2 and 5 (8 each), with zero recorded events in Tasks 4 and 8. One task could contain more than one error event, so error counts are not equivalent to task failures. Appendix Table C1 reports error categories, Appendix Table C2 provides task-level counts, and Appendix Table C3 summarizes qualitative observations. Figure 4 juxtaposes assistance and error events by task.

Mean total time to complete the full 12-task session was 333.29 s per participant (median = 297 s; range = 275–423 s; Appendix Table B2). The study archive does not retain separate time-on-task values for each task, so timing is interpreted only at the participant-session level.

The 12 SUS respondents were a separate student group from the seven task-based participants. They were 21–23 years old, with nine men and three women; reported study backgrounds included Visual Communication Design (4), Informatics (2), Civil Engineering (1), Electronics Engineering (1), Informatics Education (1), and three unspecified study programs. SUS scores ranged from 67.50 to 95.00, with a mean of 81.67 (SD = 9.25), median of 81.25, and 95% CI [75.79, 87.55]. Eleven respondents met the study’s descriptive Acceptable threshold and one was classified as Marginal.

The favorable overall SUS score coexisted with assistance during more than one quarter of the observed task executions. This divergence is the central human-centered finding: perceived ease and successful eventual completion did not eliminate first-use discoverability friction. Qualitative observations were consistent with the task counts: participants asked how to find Studio, My Projects, and Logout, questioned prompt-language flexibility and style controls, and sometimes required direction when changing model visibility.

The two expert instruments are reported separately because implementation conformity and generated-model quality are not interchangeable outcomes (Table 3). The system instrument contained 20 five-point items and produced 81/100 (mean 4.05/5). The output-quality instrument contained 10 five-point items and produced 35/50 (mean 3.50/5). Complete item-level scores are provided in Appendix Table D1 (system implementation) and Appendix Table D2 (generated-model quality).

For system implementation, the highest ratings (5/5) were assigned to system-function conformity, generation-flow conformity, and prompt enhancement; 3D output and interface clarity were lower at 3/5. For generated-model quality, GLB output suitability received 5/5; detail, surface quality, and suitability as a 3D asset received 4/5; geometry, structural completeness, shape accuracy, multiview consistency, visual conformity, and prompt/image correspondence received 3/5. The pattern supports an operational prototype with usable export while locating the principal quality constraints in geometry, structural fidelity, and cross-view consistency. Because one validator completed both instruments, no inter-rater reliability statistic can be established.

For RQ1, the results show that established generative components can be assembled into a browser-accessible workflow without proposing a new generative model. The contribution is architectural and interactional: the system converts heterogeneous stages into a coherent sequence of input, queued execution, status feedback, persistent output, inspection, and download. This application layer complements diffusion synthesis [4], latent diffusion [5], vision–language alignment [6], DreamFusion [7], Magic3D [8], ProlificDreamer [9], visual-prompt conditioning [10], and image-conditioned reconstruction [11], [12], [13], [14], [15], [16]. It also operationalizes human-centered guidance on reliable control [1], feedback and correction [2], and AI-specific design challenges [3].

The application-level contribution is more directly comparable to systems such as Codeless3D [17] and MagicCraft [18], which likewise aim to reduce specialist barriers to 3D creation. The present study adds a specific claim–evidence structure for a web-based generative pipeline: orchestration reliability, file readability, deployment turnaround, observed independent use, perceived usability, and output quality are measured separately. This separation is important because a visually plausible model can still be embedded in a poor workflow, and a technically reliable workflow can still produce geometrically imperfect assets.

Educationally, this separation is useful because technology-enhanced-learning applications require both workable tools and understandable interaction. A student authoring activity may be disrupted if users cannot find the generation workspace or retrieve their results, even when the generative model is technically capable. Conversely, easy navigation does not guarantee that a generated asset has sufficient geometry for a learning task. The present study therefore motivates, but does not demonstrate, future classroom evaluation in which interaction measures are paired with learning or project outcomes.

For RQ2, the omnibus difference among the observed turnaround distributions is best interpreted as a deployment signal that merits profiling, not as evidence of an intrinsic speed hierarchy among generation modes. ControlNet introduces additional conditioning operations and multi-image reconstruction processes additional views, making longer or more variable workflows plausible, but the retained processing_time also includes monitoring, retrieval, and storage transfer. Hardware load and input complexity were not controlled. A statistically significant omnibus result therefore cannot isolate generation mode as the causal factor.

Cross-viewer opening in the integrated viewer, Blender, and 3dviewer.net strengthens practical relevance because exported assets are not locked to the originating web interface. Nevertheless, the evidence is deliberately termed cross-viewer GLB compatibility. It demonstrates opening/readability, not complete interoperability. Future compatibility testing should add GLTF/GLB schema validation, material and texture preservation, transform checks, mesh-manifold or integrity tests, file-size reporting, and repeated import–export round trips.

For RQ3, the combined task and SUS evidence reveals a discrepancy between overall perceptions and observed behavior. Mean SUS was favorable at 81.67, and all specified tasks were eventually completed; however, 23 of 84 executions required assistance. The task-level data show that this assistance was concentrated in navigation and result-management steps. Entering Studio and returning to a generated model through My Projects each required assistance for five of seven participants, while making a model public and logging out each required assistance for three. The evidence therefore points to discoverability rather than broad functional failure.

The 48 recorded error events provide further detail. Input and navigation accounted for 44 of 48 events, whereas interpretation and procedural errors were uncommon. Interface redesign should therefore prioritize information scent and primary-action visibility before adding new features: Studio and My Projects should be prominent persistent destinations; Gallery and visibility state should use explicit labels; logout should be consistently located; prompt input should state supported languages and examples; style controls should explain their effect; and job status should communicate expected duration, queued/processing state, completion, and recoverable errors. These recommendations are directly tied to the observed error and assistance pattern rather than being generic interface advice.

The difference between task observation and SUS also illustrates why global questionnaires should not replace behavioral evaluation. SUS captures overall perceived usability [31], while its interpretation and accumulated evidence are discussed more broadly in [32]. A favorable global score can therefore coexist with assistance on specific first-use tasks. Conversely, the task study’s seven-person formative sample is not large enough to estimate population-level incidence. The two methods are therefore complementary: task observation localizes friction; SUS summarizes perception in a separate sample.

The expert output score of 3.50/5 qualifies the favorable functional and SUS findings. GLB suitability was strong, but geometry, structural completeness, shape accuracy, multiview consistency, and correspondence to prompts or images were only moderate. Related reconstruction literature addresses relevant geometric and appearance challenges through neural surface reconstruction [33], texture fields [34], volume-rendered implicit surfaces [35], and generative radiance fields [36]. The present result supports the system as an accessible research prototype and authoring aid, not as evidence of state-of-the-art reconstruction fidelity or production-ready assets without further editing.

Three transferable implications follow. First, studies of interactive generative media should report orchestration and interaction outcomes alongside model outputs because users experience the entire workflow. Second, eventual completion and global perceived usability should be separated from independent completion, assistance, and discoverability. Third, output portability should be assessed independently from output fidelity. These principles provide a reusable multidimensional evaluation framework but do not establish computational reproducibility.

Several limitations constrain generalization. Timing data include only completed jobs, use unbalanced mode counts, combine heterogeneous prompts and images, and lack preserved hardware telemetry, package versions, random seeds, dependency locks, and system-load measurements. Task testing involved seven convenience-sampled students; the separate SUS sample contained 12 students; prior 3D/generative-AI experience was not retained; and no baseline interface or longitudinal use was tested. The task archive provides task-level completion and error counts but not per-task time or a fully reproducible facilitator-intervention threshold. Functional testing includes a complete retained 14-scenario matrix, but the individual tester identity and immutable build identifier are not preserved. Only one expert rated implementation and 3D quality, so inter-rater reliability cannot be estimated. The study did not include production load, security/stress testing, a complete failure-rate denominator, objective geometric ground truth, blinded perceptual comparison, or educational outcomes.

The available ethics documentation is limited. The study was conducted as a non-sensitive, non-interventional software-usability evaluation within the undergraduate research process. Participation was voluntary and informed consent was obtained before task-based testing or SUS participation. According to the institutional requirements recorded for the study, formal ethics committee clearance was not required. The retained research record does not include a prospective ethics-exemption letter or identifier, so this determination cannot be independently substantiated from the archived documentation.

Future work should preserve containerized environment manifests and hardware telemetry, balance and randomize mode-specific workloads, retain failed and cancelled jobs, and profile each pipeline stage. Human-centered evaluation should recruit broader user groups, document prior experience, define assistance thresholds prospectively, retain per-task timing, compare revised interfaces, and examine learnability over repeated sessions. Output evaluation should add multiple blinded raters, inter-rater reliability, reference-based geometry metrics where ground truth exists, and perceptual comparison against alternative workflows. Educational extensions should test whether generated 3D assets improve specific learning, design, or communication outcomes rather than assuming benefit from usability alone.

Not applicable. The article does not report identifiable participant images or confidential personal records.

This article is derived from the first author's undergraduate thesis and is not under consideration elsewhere.