When AI Restores a Photo, Is It Preserving the Past or Inventing It?

A restored photograph can look clearer while becoming harder to verify. Generative face-restoration research shows why: when an image has lost information, a model can supply plausible detail from patterns it learned during training. The result may resemble the person more closely, yet still contain features the surviving photograph cannot establish. CodeFormer’s research makes that tension explicit.

That leaves a practical question for anyone restoring a family collection: what should the next viewer be able to trust? A useful answer begins with separating the image we inherited from the reconstruction we choose to make. Both can have a place, provided the relationship between them stays visible.

What the model is actually recovering

In its 2022 NeurIPS paper, CodeFormer uses a learned codebook and a neural network to reconstruct degraded faces. The codebook stores learned representations used to generate a face. The system also offers a control that changes the balance between visual quality and fidelity to the input. Its authors report remaining problems with accessories and side-facing portraits. These are limitations of that system, not evidence that every restoration tool fails in the same way.

A broader explanation comes from Yochai Blau and Tomer Michaeli’s peer-reviewed perception-distortion research. Their analysis distinguishes an output’s resemblance to natural images from its similarity to the particular original. Near an optimal performance boundary, improving one can come at the expense of the other. The result does not mean that every sharper photograph is less accurate, or that better algorithms cannot improve both measures.

For a photograph’s historical use, the implication is narrower and more useful: a convincing appearance alone cannot settle whether a reconstructed detail was present. A quality score answers a technical evaluation question. Establishing what someone actually wore, or what a damaged part of a photograph contained, requires evidence for that particular detail.

The case for restoring something valuable

The stakes become concrete in Valencia, Spain. Following the October 2024 floods, volunteers working with Universitat Politècnica de València cleaned and digitized damaged family photographs. A 2026 AAAI demonstration paper describes Rebrot, an AI system developed within that recovery effort to detect damaged regions and reconstruct missing content.

The researchers also describe a revealing problem: the general restoration process could subtly alter faces. They therefore created a separate facial-processing path intended to preserve identity. The paper documents that design choice, but its three pages provide no numerical evaluation results demonstrating historical accuracy. Its reference to more than 200,000 affected photographs describes the recovery project’s scale, not a verified count of successful AI restorations.

This is a serious reason to develop restoration tools. Making a damaged family image usable again is a worthwhile goal. The same example also shows why a promise to preserve someone’s likeness deserves careful checking. A humanitarian purpose and a technically plausible result do not, by themselves, verify every reconstructed feature.

What memory research can actually tell us

A peer-reviewed CHI 2025 study by Pat Pataranutaporn and colleagues examined 200 U.S. participants’ recall of recently viewed images. Participants saw 24 original images, completed a two-minute filler task, and then saw unedited images or one of three AI-altered conditions. They answered questions about the originals.

In the corrected published paper, the mean percentage of incorrect answers on the 12 targeted images was 18.9% in the control condition and 31.4% after AI-edited still images. The researchers classified those errors as false memories. These percentages describe answers within the experiment, not the proportion of people who acquired a false autobiographical memory.

The distinction matters. This study tested short-term recall after changes to image content. It did not test repairing family photographs, memories accumulated over decades, or what happens after years of viewing a restored portrait. It also lacked a matched manual-edit condition, so it cannot establish that AI edits are uniquely harmful compared with equivalent human-made alterations.

Another peer-reviewed experiment provides a useful check on that assumption. In Gillian Murphy and colleagues’ 2023 PLOS ONE study, 436 participants encountered descriptions of real and fictitious film remakes, sometimes accompanied by clips. False-memory rates did not differ significantly between deepfake and text presentations for any of the fictitious films. The study’s broad memory measure included reporting having heard of a film. This result does not establish that deepfakes are harmless; it cautions against treating the technology alone as the explanation.

Together, these studies justify taking altered visual evidence seriously while leaving the central family-restoration question open. Neither establishes how ordinary restoration changes long-standing memories of a relative.

More evidence can improve the reconstruction

There is also technical work aimed at reducing unsupported detail. In a peer-reviewed ICLR 2025 paper, Keda Tao and colleagues introduce MGFR, which guides face restoration with attribute descriptions, reference photographs, and identity information. Their comparisons show how references can help reconstruct details such as wrinkles and moles that are difficult to recover from a degraded input alone.

Those results need context. Some comparison methods did not receive reference photographs, so those comparisons cannot isolate the advantage of the model’s architecture. The same paper demonstrates controls that can add accessories or change identity. Guidance helps determine the result; it does not automatically make the result historically correct.

The practical inference is to ask where each instruction came from. A feature supported by another reliable photograph has a different evidentiary basis from one requested because it looks appealing. When the available records cannot resolve a detail, leaving it uncertain is a defensible editorial choice.

What the next viewer should know

Preservation professionals already distinguish faithful recording from interpretive intervention. The Federal Agencies Digital Guidelines Initiative’s 2023 guidance says digitization should document an object as it appears at capture, rather than recreate a conjectured earlier condition. It also calls for documentation of processing and relationships between different file versions. This is authoritative professional guidance, not a peer-reviewed experiment about memory.

The American Institute for Conservation’s guidelines likewise call for loss compensation to be documented, detectable, reversible, and consistent with known characteristics. These conservation principles offer a useful basis for deciding how to handle digital reconstructions. They do not establish a tested method for preventing memory distortion.

For a family collection, the following approach applies those principles:

• Keep the physical photograph and a faithful digital capture. Save the reconstruction separately.

• Record the tool, settings, date, and substantial changes, along with any reference photographs used.

• Label generated or uncertain details specifically. Explain whether work removed damage, supplied color, reconstructed facial features, or added movement.

• Share the original alongside the reconstruction when practical, so viewers can examine the evidence themselves.

These are recommendations for transparency, not proven safeguards against false memories. Their value is that they preserve the ability to question an edit and revisit it as better evidence becomes available.

AI can help make a damaged photograph accessible again. It can also fill gaps with choices that the original cannot confirm. Responsible restoration should leave room for both the pleasure of a clearer image and an honest account of how it was made. When that photograph reaches the next generation, they should inherit the surviving evidence, the reconstruction, and enough context to tell them apart.

Sources

The research cited above comes from peer-reviewed conference proceedings and a peer-reviewed journal. The two preservation documents are identified separately as professional guidance.

• Blau, Y., and Michaeli, T. (2018). The Perception-Distortion Tradeoff. CVPR, 6228–6237.

• Zhou, S., Chan, K. C. K., Li, C., and Loy, C. C. (2022). Towards Robust Blind Face Restoration with Codebook Lookup Transformer. NeurIPS 35.

• Tao, K., Gu, J., Zhang, Y., Wang, X., and Cheng, N. (2025). Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model. ICLR.

• Pataranutaporn, P., et al. (2025). Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection. CHI. DOI: 10.1145/3706598.3713697. Corrected version of record, June 17, 2025.

• Murphy, G., Ching, D., Twomey, J., and Linehan, C. (2023). Face/Off: Changing the face of movies with deepfakes. PLOS ONE, 18(7), e0287503.

• Galvan Suazo, J. D., Albert Bonet, H., Monserrat-Aranda, C., and Ferri, C. (2026). AI for Memory Preservation: Automated Restoration of Photographs Damaged by Floods. AAAI, 40(48), 41691–41693. Demonstration paper.

• Professional guidance: FADGI Technical Guidelines for Digitizing Cultural Heritage Materials, third edition (2023); American Institute for Conservation Code of Ethics and Guidelines for Practice, especially guidelines 23–28.