Page 85 - AC-1-2

Arts & Communication Digital restoration with generative AI

3.4. Stable diffusion terminology and technique

Within the Stable Diffusion framework, several
terminologies define the nuanced processes integral to
restoration. The term “inference steps” pertains to the
methodical enhancement of the image, with each step
leading to further refinement. The stabilization of this
enhancement is embodied in the “diffusion” process. The
use of “guidance” signifies the deployment of reference
imagery to bolster accuracy, while “prompt strength”
dictates the degree of initial input into the system. The
resultant image’s clarity and authenticity are encompassed
under “image quality.” Of paramount significance to
this research, the “inpainting” procedure encompasses
the act of designating specific regions of the image for
modification or the infusion of new details, as articulated
Figure 1. Antoine François Callet, Achilles Dragging Hector’s Body Past the
Walls of Troy (1784–1785), Louvre, Paris. Source: Photo by the authors on in Stable Diffusion 2.1 (2022).
November 16, 2022.
3.5. Stable diffusion prompting procedures
In undertaking the restoration process, the Stable Diffusion
framework was initialized using a random seed and
tailored to accommodate an image dimension of 768 × 768
pixels. The restoration commenced with an initial series
of 25 inference steps, followed by additional sequences of
25 steps each, all meticulously evaluated for their visual
consistency and quality. Within this process, a guidance
scale of 7.5 was consistently adhered to. The prompt
strength oscillated between 0.25 and 0.5, an adjustment
designed to ensure an equilibrium between authenticity and
revitalization during the restoration. Aiming to mirror the
integrity of the original piece, the image quality parameter
was anchored at 75. Furthermore, during the inpainting
process, the prompt strength was elevated to 0.7, thereby
broadening the spectrum of potential restoration nuances.
The methodological exactitude underpinning this
research elucidates the complementary relationship
between time-honored art conservation methods and the
rapidly evolving domain of AI. Following this methodology
exposition, the forthcoming section will transition to an
analysis of the findings, placing a pronounced emphasis on
the role of generative AI in the rehabilitation of imperiled
or vanished elements within the vast repository of cultural
heritage.
Figure 2. Pietro Antonio Martini, View of the Salon 1785, 1785, engraving.
Metropolitan Museum of Art, New York (detail of Antoine François 4. Results and discussion
Callet, Achilles Dragging Hector’s Body Past the Walls of Troy (1784–1785),
Louvre, Paris). Source: Photo by the authors on November 16, 2022. During the initial phase of restoration, a meticulous manual
technique was employed, targeting a specific segment of
their neighboring regions, thereby curbing any visual the original painting (Figure 3). Intensive color masking
discrepancies. To further enhance visual cohesiveness, strategies were implemented, complemented by the
supplemental layers of color masks were superimposed onto introduction of noise and caustic effects to accentuate select
the background, with specific emphasis on areas proximal attributes. Despite these efforts, the outcome rendered only
to the female figure and the depicted limbs. a slight revitalization of the deteriorated segments of the

Volume 1 Issue 2 (2023) 4 https://doi.org/10.36922/ac.1793