Od rekonstrukcji do kreacji: generatywna AI w pracy z obrazem archiwalnym [Panel III]

Value of the Information & Strength of the Argument

The presentations provide substantial value by bridging technical AI capabilities with historical methodology. Wilkowski adeptly deconstructs the colorization process into layers, illustrating how technical decisions affect historical interpretation. He argues effectively that training data biases lead to ‘presentification’ and ‘ahistorical’ outputs, using vivid examples like the homogenized dogs in a colorized 1936 photo. Zając’s empirical account shows the evolution of AI tools and their practical limitations, demonstrating that while fully automatic results may be flawed, user-guided refinement can yield compelling outcomes. The argumentation is coherent, acknowledging both the potential and the dangers of these technologies, and advocating for ethical guidelines and critical literacy. The discussion is well-structured, with each speaker complementing the other’s perspective, and the eventual Q&A likely underlines the urgency of establishing archival standards.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high in the sense that the speakers refer to specific models (DeOldify, ChromaGAN, Unicolor), datasets (ImageNet), and theoretical concepts (e.g., ‘presentification’, ‘ahistorical’), grounding their claims in observable phenomena. However, the lack of formal citations—sources are mentioned informally without full references—reduces verifiability. The title accurately reflects the content, as the panel indeed moves from reconstruction to creation, questioning the boundary between restoration and fabrication. The adéquation between the title and the delivered talks is strong, with both speakers directly addressing the implications for archival practice.

The title accurately reflects the session's focus on generative AI for archival images, covering both reconstruction and creation concepts.

8/10

The presentation features two specialists in digital history and photography, offering technically detailed explanations and practical demonstrations. Sources are primarily model names and general references, but the discussion is grounded in known research and practical experience. The reliability is high for the expertise conveyed, though verification of specific claims would require external literature.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:45Moderator Mariusz Żuławnik introduces the panel and speakers.
• 05:00Marcin Wilkowski begins his talk on levels of processing in AI colorization.
• 15:00Wilkowski explains the technical limitations of AI colorization, including training data bias.
• 25:00Wilkowski discusses semantic transformations like 'presentification' and 'ahistorical' outcomes.
• 35:00Mariusz Zając starts his presentation on using AI to restore archival films and photos.
• 50:00Zając shows examples of AI-enhanced reconstructions from pre-war Warsaw.
• 1:00:00Discussion begins on the risks of AI-generated forgeries in archives and potential mitigations.

• DeOldify — Zając mentions using DeOldify in 2020 for colorization, leading to his project Historical Colors.
• ChromaGAN — Wilkowski refers to this model as an example of automatic colorization based on object recognition.
• Unicolor — Wilkowski cites this as a mixed-method approach for user-guided colorization.
• ImageNet — Wilkowski mentions this dataset as a contemporary source that introduces bias in historical colorization.

• ImageNet — Mentioned as the training dataset with contemporary images, supporting the claim about bias.