Part 2: Generative AI in Photo Restoration
A hobbyist’s guide to ChatGPT/Gemini AI image engines without losing resolution or authenticity
Original "Ultima Thule" daguerreotype of Edgar Allen Poe
Why This Article Exists
If you restore photographs seriously -even if not professionally -you are operating in a rapidly shifting environment. AI-powered image tools are now widely available, highly capable, and heavily promoted across social media, forums, and YouTube.
Many hobbyists understandably ask:
“If AI can do this so well now, why not just use it?”
The honest answer is sometimes you can but understanding when, how, and at what cost matters far more than most tutorials acknowledge.
This guide is written for advanced hobbyists, family historians, and skilled amateurs who may already use tools like Lightroom, Photoshop, and Topaz, and who want to incorporate generative AI without accidentally discarding resolution, altering identity, or misrepresenting the original photograph.
Restorative AI vs Generative AI (And Why It Matters)
Much of the confusion around AI in photo restoration comes from treating all AI tools as interchangeable.
Restorative AI
Restorative AI tools analyze existing pixels and attempt to refine them—reducing noise, improving sharpness, or clarifying edges—while preserving the underlying geometry and relationships already present in the image.
Examples include:
Lightroom/Photoshop – e.g. repair, denoise and sharpen
Topaz Photo AI – e.g. upscaling, sharpening and denoise
These tools work with your image data - repairing and replacing missing pixels or multiplying existing pixels to upscale.
Generative AI (ChatGPT/Gemini-style image engines)
Generative AI is pixel creative. It synthesizes new pixels when information is missing or ambiguous. They do not enhance your pixels; they replace them with plausible alternatives.
This is not inherently bad—but it is a fundamentally different operation. For hobbyists, the key takeaway is simple:
Generative AI creates convincing images. It is not true to a photo’s historical accuracy.
Generative AI with a Basic Prompt
“Restore this old photo to look as if it were taken today with a high-resolution modern digital camera. Refine and sharpen the background and make it realistic.”
These results give a good ‘rendition’ of the original photo with only one pass but upon closer inspection many of the details of the individuals and background have been recreated to give a good proximity to the original but not a precise replication. Many newcomers to restoration will be very pleased with these results. For print sizes below 8”x10” they may be acceptable.
The Resolution Trap - What Hobbyists Most Often Miss
One of the most common mistakes hobbyists make when experimenting with generative AI is using it as a first-pass restoration tool on a high-resolution scan.
When a 20–40 MP scan is processed through a generative system that outputs a smaller image, the original scan data has effectively been replaced. Upscaling afterward cannot recover what was lost—it can only invent new detail.
For sharing online or creating a quick family keepsake, this may be acceptable.
For preserving a meaningful photograph, it is often an unintended compromise
A practical rule of thumb for hobbyists:
Restorative AI = improves your pixels
Generative AI = replaces your pixels
Hallucination Isn’t Random - It’s Structural
AI “hallucination” often gets framed as an occasional failure. In restoration work, it’s better understood as a design feature:
Faces receive the highest priority (e.g. eyes, nose, mouth, expression)
Clothing and backgrounds are treated as context and are often generalized
Fine textures (fabric weave, film grain, paper tooth) are frequently simplified
A clean, modern aesthetic is favored
This is why AI restorations often produce striking faces paired with slightly generic clothing or backgrounds. For casual use, this may be perfectly fine. For historically meaningful photos, it deserves careful scrutiny. Look at the above example on the right, AI created new trees in the background.
Prompting Discipline - Why “One Careful Pass” Beats Many
Vague prompts such as “restore this photo” give AI systems wide creative latitude. In response, many users run multiple AI passes trying to “fix” issues they notice.
Ironically, this often makes things worse.
Each generative pass treats the previous output as ground truth, gradually drifting farther from the original photograph.
If a single, carefully constrained prompt does not produce an acceptable result, it is usually a sign that manual restoration - not more AI -is the better next step.
Canonical ‘One-Pass’ Prompts for Hobbyist Use
The following prompts are intentionally detailed. They are designed to limit identity drift, avoid beautification, and reduce invented detail.
These should be used once per image, not iteratively. If it’s not acceptable, switch to manual methods rather than looping AI.
These prompts acknowledge tradeoffs and are intended for one-pass use only. A caution here - some clients may find these do not “modernize” the restorations enough.
Single Portrait Prompt
“Restore this photograph carefully while preserving the subject’s exact facial structure, proportions, expression, and identity. Do not alter bone structure, eye shape, nose shape, mouth shape, or overall likeness. Repair visible damage such as scratches, dust, stains, fading, or minor blur without inventing new details. Improve clarity and sharpness only where supported by the original image. Maintain natural skin texture and age-appropriate detail; do not beautify, smooth, or stylize the face. Preserve original lighting direction and shadow relationships. Clothing and background should be enhanced conservatively, without adding texture, patterns, or details that were not clearly present. The result should look like a clean, high-quality scan of the original photograph, not a modern reimagining.”
Group Portrait Prompt
“Restore this group photograph while preserving each individual’s facial structure, proportions, expression, and relative appearance exactly as shown. Do not modify any person to look younger, sharper, or more prominent than others. Repair damage such as scratches, dust, stains, and fading while maintaining consistent detail and contrast across all subjects. Avoid enhancing one face more than another. Do not invent facial details, hair texture, clothing features, or background elements. Preserve original spatial relationships, posture, and scale. The final image should appear as a faithful, cleaned version of the original group photograph, suitable for historical or family reference.”
Scene/Landscape Prompt
“Restore this photograph by repairing damage such as scratches, dust, stains, fading, and uneven exposure while preserving the original composition and geometry. Improve clarity and tonal balance without adding, removing, or reshaping objects, structures, or natural features. Maintain realistic textures appropriate to the era of the photograph; do not introduce modern sharpness, stylization, or dramatic contrast. Preserve the original lighting conditions and atmosphere. The result should resemble a high-quality restoration of the original scene, not a modern reinterpretation.”
Colorization Prompt (Apply After Restoration)
“Apply colorization conservatively and in a historically appropriate manner based on the era suggested by the photograph. Preserve all facial features, textures, and structural details exactly as they appear in the restored image. Use natural, subdued colors consistent with period clothing, skin tones, and materials. Avoid modern color palettes, excessive saturation, or stylistic effects.”
Colorization should always be treated as interpretive, not factual. Even when historically informed, AI-based colorization reflects probability, not certainty.
Case Study: The Edgar Allan Poe Prompt (Why Results Vary)
To better understand how generative AI allocates detail, the same historical photograph of Edgar Allan Poe was processed using identical prompts across different AI systems.
Verbatim Prompt Used (As Run)
“Restore and enhance this photo as if it were captured today with a high-end modern camera. Improve overall sharpness and fine detail without changing the subject's identity or features. Enhance colors to look natural and vibrant (not oversaturated), correct white balance, and gently increase contrast and dynamic range. Reduce noise, fix blur or softness, and refine edges for a crisp, realistic look. Preserve authentic skin tones and textures, avoiding any artificial smoothing or stylization. The final image should feel clean, high-resolution, and true to the original scene, just clearer and more lifelike.”
What matters for hobbyists is not which version “looks better,” but where each system chose to spend detail.
Across outputs:
Facial features were emphasized
Clothing texture was simplified
Backgrounds were generalized
This behavior is consistent and predictable - and explains why AI restorations often feel simultaneously impressive and subtly incomplete.
How To Evaluate AI Restoration Outputs (A Hobbyist’s Checklist)
Before accepting any AI-generated restoration, review the result at 100% zoom using the checklist below. This step alone will eliminate most unintentional damage caused by generative tools.
Identity & geometry
Are eyes, nose, mouth shapes identical to the original (not ‘improved’)?
Did expression shift (more alert, more cheerful, less tired)?
Did the head shape or jawline subtly change?
Texture integrity
Did skin texture become plasticky or uniformly smooth?
Did hair become a generic ‘painted’ texture?
Did fabric gain a suspiciously uniform weave or lose era-specific characteristics?
Historical fidelity
Are uniform details, buttons, insignia, jewelry, or lace patterns still truthful?
Did the background become a modern gradient or studio look?
Output viability
Is the file resolution adequate for the client’s intended print size?
If you must upscale, are you upscaling the whole image or only a repaired patch?
A ‘Sandwich’ Hybrid Workflow That Keeps You in Control
If you want speed without sacrificing integrity, a hybrid workflow offers the best balance:
1) Preserve the master
Keep your high-resolution scan (TIFF/PSD) as the digital negative
2) Do safe global work first
Lightroom or Photoshop’s Camera Raw for tonal corrections
Topaz, Photoshop or Lightroom for denoise/sharpen (restorative AI only)
3) Use generative AI surgically
Only for missing data or non-critical regions, preferably as patches
4) Reintegrate texture
Bring back grain/paper texture from the scan (overlay/soft light) as needed
5) Upscale last
Decide your intended use digital or print and size accordingly e.g. an 8”x10” print upscale to at least 2400x3000 px.
This approach preserves the original photograph as the source of truth while still benefiting from modern AI tools.
Practical Patching Rules (What I Would Actually Do)
Never send the full-res master to a generative engine if it will come back smaller
If generative repair is required, isolate the tear, corner, or background and keep faces outside the selection
Treat tools such as Photoshop’s Generative Fill (and future inpainting-optimized versions of Gemini or ChatGPT) as patch tools, not restoration engines
Always blend generated patches with masks; avoid hard edges
After blending, reintroduce original grain globally so the patch doesn’t look artificially clean
Clearly disclose when colorization is interpretive, especially for uniforms and historically sensitive subjects
Client Communication - A Simple Disclosure That Builds Trust
If you use generative AI in any part of a paid restoration, consider a short, plain-English disclosure such as:
“This restoration includes careful manual work and may include limited AI-assisted repair for missing or heavily damaged areas. All work is performed with the goal of preserving the original likeness and historical integrity. Colorization, when requested, is an interpretive rendering.”
This isn’t legalese. It’s trust-building. It also clearly differentiates your work from casual social media restorations.
When AI Is ‘Good Enough’ - and When It Isn’t
In reality, some people want a fast, inexpensive, visually pleasing image for sharing. In those cases, a disciplined one-pass AI restoration using the prompts outlined earlier may be entirely appropriate.
However, for:
Archival preservation
Large-format prints
Historically significant subjects
Family heirlooms meant to last
Generative AI should be limited to patching missing data and should never replace the original scan as the primary source of truth.
Final Thoughts
The professional advantage is restraint.
Generative AI is not the enemy and not the savior. It is a powerful tool that can reduce labor when used with discipline.
Your competitive advantage as a restorer is not access to AI. Everyone has that now. Your advantage is judgment - knowing when to use it, where to confine it, and when to stop.
Appendix: Comparing AI Restorations
Edgar Allan Poe Case Study
(What Each Model Did Differently)
I Asked Both AI Generators to Review and Critique Each Other’s Image
To understand how generative AI systems behave in restoration scenarios, the same historical photograph of Edgar Allan Poe was processed using the same prompt across two different AI image engines. The goal was not to determine a “winner,” but to observe how each system allocates detail and makes tradeoffs.
The prompt used was identical in both cases and asked for a careful restoration that preserved identity while improving clarity and realism.
What follows is a focused, side-by-side analysis of the results from the creators’ themselves. What is interesting is the bias they show in favoring their versions over the competitor.
Here is Gemini’s review:
ChatGPT Image
This image leans heavily into the "Modern Camera" part of the prompt. It has the distinct crispness of a high-megapixel digital sensor.
Strengths
Micro-Contrast - The texture in the skin and the fabric of the coat is incredibly sharp. You can see individual pores and thread weaves.
Intensity - It captured an intense, almost piercing stare. The eyes are very clear and reflective.
Dynamic Range - It pushed the lighting to be very dramatic, with deep blacks in the coat and bright highlights on the forehead.
Weaknesses
It feels slightly hyper-real or "digital." The sharpness is so aggressive that it crosses into the "Uncanny Valley"- it looks less like a restored photo of a man from the 1840s and more like a high-end CGI render or a modern actor in makeup.
The eyes, in particular, feel a bit "wide" and modernized.
Gemini Image
My generation tends to lean more towards "Restoration & Artistry." It usually prioritizes preserving the "soul" of the original daguerreotype rather than inventing new high-frequency detail.
Strengths
Atmosphere - My version likely retained more of the soft, melancholic lighting typical of 19th-century portraiture, even while colorizing it.
Naturalism -The skin tones and transitions are usually smoother. It often looks more like a painting or a soft-focus portrait than a harsh digital photo.
Fidelity - It often tries to keep the facial structure softer, which can feel truer to how we remember Poe, rather than sharpening him into a modern aesthetic.
Weaknesses
In a direct "sharpness" contest, my version often loses to ChatGPT. It might smooth over details (like the texture of the cravat or the grit of the mustache) that the other model explicitly enhances.
If you are looking for "grit," my version can sometimes feel too "clean."
Here is ChatGPT’s review:
ChatGPT Image
Strengths
Facial proportions are preserved almost exactly:
Eye spacing
Nose length
Brow structure
Jaw width
Expression remains somber and tense, matching the original
Skin texture shows age-appropriate unevenness, not cosmetic smoothing
Weaknesses
Slight softening around the temples and under-eye areas
Micro-wrinkle detail is inferred, not recovered
Gemini Image
Strengths
Higher perceived sharpness in eyes and nose
Stronger local contrast gives immediate “pop”
Weaknesses
Subtle reshaping:
Eyes appear fractionally larger
Cheeks slightly fuller
Expression drifts toward “alert” rather than “haunted”
Skin texture begins to look manufactured, not photographic
What This Comparison Actually Shows
Taken together, these two restorations demonstrate that generative AI models do not fail or succeed in the same ways. Each allocates attention differently, emphasizing certain visual priorities while simplifying others. Facial identity may be preserved in both cases, yet texture, fabric detail, background treatment, and overall aesthetic intent diverge in predictable ways.
The value of this comparison is not in deciding which result is “better,” but in understanding where interpretation occurs and why. For hobbyists, this reinforces an essential lesson: AI restorations must be evaluated region by region, not judged by first impression. The more clearly you can identify what was preserved, what was generalized, and what was invented, the more responsibly you can decide whether an AI-generated result is acceptable for the photograph’s purpose.
