Page 86 - AC-1-2

Arts & Communication Digital restoration with generative AI

artwork. The resultant image conspicuously underscored crucial role of human oversight in ensuring accuracy and
the constraints tied to the nuances of human craftsmanship authenticity.
and interpretation. Venturing further, the study experimented with an
In the ensuing phase, AI mechanisms were employed, integrated approach, harmonizing manual restoration
leveraging the capabilities of Stable Diffusion. The techniques with AI functionalities (Figure 5). A precise
parameters set for this endeavor included a prompt strength descriptive prompt was input into the AI model, elaborating
of 50%, a guidance scale of 7.5, and a commencement of 25 on various facets and elements intrinsic to the painting. This
inference steps (Figure 4). Intriguingly, the AI exhibited was complemented by a host of technical determinants,
a propensity for misinterpretation: rather than faithfully including seed values and specified dimensions. In this
restoring the damaged sections, the AI reconstructed phase, the inpainting function was employed judiciously
them as entirely distinct celestial entities. Such an outcome to rectify minor imperfections. The parameters deployed
underscores the challenges AI encounters in deciphering were comprehensive: encompassing a detailed description
and preserving the inherent artistic intent, echoing of the painting’s elements, a designated seed, explicit
the insights propounded by Yu et al. (2021) about the dimensions, and other technicalities such as inference
steps and the guidance scale. Notably, the amalgamation
of manual expertise with AI acumen culminated in an
outcome that bore a close resemblance to the original
artwork. However, this outcome was not without minor
inconsistencies, especially in terms of color fidelity and the
visibility of certain superficial imperfections.

An exhaustive exploration into the efficacy of
inpainting was undertaken, marked by multiple iterations
to determine optimal outcomes (Figure 6). Following
the initial two rounds, a conspicuous reduction in visible
imperfections was observed. However, this came at
the detriment of the integrity of color representation,
especially affecting the nuanced details of Achilles’ hands
and facial contours. Proceeding to the fifth round, an
almost complete obliteration of scratches was achieved,
yet this was accompanied by an altered color gradient. By
Figure 3. Manual painting and recolorization. Source: Photo by the the ninth round, a paradoxical outcome emerged: while
authors of their own portrayal of Achilles Dragging Hector’s Body Past the surface imperfections were entirely effaced, there was a
Walls of Troy on May 22, 2023. discernible compromise in the granularity of intricate

Figure 4. Artificial intelligence-based restoration. Source: Photo by the Figure 5. Confluence of manual and artificial intelligence techniques.
authors of their own portrayal of Achilles Dragging Hector’s Body Past the Source: Photo by the authors of their own portrayal of Achilles Dragging
Walls of Troy on May 22, 2023. Hector’s Body Past the Walls of Troy on May 22, 2023.

Volume 1 Issue 2 (2023) 5 https://doi.org/10.36922/ac.1793