Is There a Perfect Prompt Alone in Image Generation? Visualizing Designerly Thinking with Stable Diffusion and ControlNet: The Case of Juicy Salif

Generative Artificial Intelligence (AI) offers novel practice-based systems for the design discipline, rapidly visualizing ideas. However, these technologies must evolve from producing unpredictable results into controllable, systematic tools. This study examines how design discourse about an existing object, translated into prompts, guides visual outputs and presents a representational fidelity framework in the search for the “perfect prompt” utilizing Stable Diffusion and the ControlNet extension. Philippe Starck’s iconic Juicy Salif design serves as the research sample. During the experiment, Starck’s original discourses and external descriptions were transformed into prompts across a three-stage process: text-only, sketch-conditioned, and reference-conditioned production. The visual outputs were analyzed using CLIPScore (semantic alignment), Learned Perceptual Image Patch Similarity (LPIPS) (perceptual similarity), and Large-scale Artificial Intelligence Open Network (LAION) Aesthetic Score (aesthetic quality). Each condition was generated with ten random seeds (ninety images) and compared with Friedman and Holm-corrected Wilcoxon tests. In the initial text-only phase, prompts derived from Starck’s original discourses achieved the highest CLIPScore, indicating that semantically dense design discourse aligns most closely with the model’s semantic representation. However, text-only outputs remained perceptually distant from the original object. Reference-image conditioning via ControlNet produced the outputs closest to the original and of the highest aesthetic quality, whereas sketch conditioning did not improve similarity, reflecting the sketch’s role as an early-stage design representation rather than the final object. Ultimately, the study contributes to the literature by proposing a controllable, metric-based systematic workflow that centers the designer as a “process manager.” This approach links initial exploration with controlled generation and supports a more reliable collaboration between designer and AI.

Authors

Institutions

Publication Details

Journal
The International Journal of Designed Objects
Published
2026-11-16
DOI
https://doi.org/10.18848/2325-1379/cgp/a164
Primary Topic
Generative Adversarial Networks and Image Synthesis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Is There a Perfect Prompt Alone in Image Generation? Visualizing Designerly Thinking with Stable Diffusion and ControlNet: The Case of Juicy Salif

Özge Kandemir, Celal Murat Kandemir, Rabia Berber, Elif Önaşçı
2 links
The International Journal of Designed Objects
Generative Adversarial Networks and Image Synthesis
article

Is There a Perfect Prompt Alone in Image Generation? Visualizing Designerly Thinking with Stable Diffusion and ControlNet: The Case of Juicy Salif

Özge Kandemir, Celal Murat Kandemir, Rabia Berber, Elif Önaşçı
article en

Abstract

Generative Artificial Intelligence (AI) offers novel practice-based systems for the design discipline, rapidly visualizing ideas. However, these technologies must evolve from producing unpredictable results into controllable, systematic tools. This study examines how design discourse about an existing object, translated into prompts, guides visual outputs and presents a representational fidelity framework in the search for the “perfect prompt” utilizing Stable Diffusion and the ControlNet extension. Philippe Starck’s iconic Juicy Salif design serves as the research sample. During the experiment, Starck’s original discourses and external descriptions were transformed into prompts across a three-stage process: text-only, sketch-conditioned, and reference-conditioned production. The visual outputs were analyzed using CLIPScore (semantic alignment), Learned Perceptual Image Patch Similarity (LPIPS) (perceptual similarity), and Large-scale Artificial Intelligence Open Network (LAION) Aesthetic Score (aesthetic quality). Each condition was generated with ten random seeds (ninety images) and compared with Friedman and Holm-corrected Wilcoxon tests. In the initial text-only phase, prompts derived from Starck’s original discourses achieved the highest CLIPScore, indicating that semantically dense design discourse aligns most closely with the model’s semantic representation. However, text-only outputs remained perceptually distant from the original object. Reference-image conditioning via ControlNet produced the outputs closest to the original and of the highest aesthetic quality, whereas sketch conditioning did not improve similarity, reflecting the sketch’s role as an early-stage design representation rather than the final object. Ultimately, the study contributes to the literature by proposing a controllable, metric-based systematic workflow that centers the designer as a “process manager.” This approach links initial exploration with controlled generation and supports a more reliable collaboration between designer and AI.

The International Journal of Designed Objects
Eskisehir Technical University (TR), Eskişehir Osmangazi University (TR)
Generative Adversarial Networks and Image Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon