My father died in the spring of 2022. Pancreatic cancer, eleven weeks from diagnosis. He was a quiet man who hated being photographed, so when we sat down to choose a picture for the funeral program, the family had perhaps fifteen photos of him from the last decade to choose from. Two were in focus. None of them looked like him.
The one we ended up using was from a wedding in 2018. He was looking off-camera at something — I think a child running past — and his mouth was caught between expressions. My mother said, that night, "It isn't him." She was right, but it was the best we had.
I spent the next eighteen months learning how identity-preserving diffusion models work — InstantID first, then PuLID, then the early Flux-Kontext line. I had a professional reason; I was building DuoPortrait. I also had a private reason. I wanted to make one good picture of my father, looking the way I remember him, even if no camera ever caught that look.
This is what I have learned about doing that carefully. I am writing it because the questions arrive in my inbox almost weekly now, and most of what is written online about "AI restoration of deceased loved ones" is either technically wrong or emotionally tone-deaf.
First, the part that nobody writes about
Before we talk about pixels: please pause before you do this.
A photograph of someone who is gone has a particular weight. It is one of the few traces left that the rest of the world recognizes. If you make a new image of that person — even a beautiful, true-feeling image — you are introducing a face into the family record that was not there yesterday. Some people in the family will find this comforting. Some will find it disorienting. A few may find it painful in a way they cannot articulate for months.
I made my father's picture and waited two weeks before showing my mother. When I did, she cried, and then she said, "Thank you. Don't make another one." I have not.
This is the rule I now give to friends who ask: make one, not a series. A single, carefully chosen image of the person, in a moment that honors who they were. Not a slideshow, not a fantasy life. Grief does not need more material; it needs one good still frame to rest on.
If you have any doubt about whether a particular family member would want this to exist — call them and ask. Do not surprise people with the face of someone they are mourning.
The American Psychological Association's guide on continuing bonds in grief is worth reading before you start, especially if the loss is recent.
The technical part
There are three different things people often mean by "AI photo restoration":
- Repair an existing damaged photograph. Scratches, fading, low resolution.
- Recreate a photograph that exists only in a small or low-quality file. A 480×640 JPEG from 2003.
- Generate a new photograph of the person in a scene that was never photographed. A grandparent holding a grandchild born after they passed.
Each one has different ethical and technical weight. I will go through them in order.
1. Repairing a damaged original
This is the easiest and the least morally complicated. The photograph already exists; you are removing scratches, dust, color shifts, and emulsion damage. The person's face is unchanged.
Modern open-source restoration is good enough that you do not need a professional service for most cases. GFPGAN and Real-ESRGAN handle 90% of family album work cleanly. You upload the scan, the model removes the damage and sharpens the face, you get the same person back at higher fidelity.
The discipline here is: restore once, save the original separately, do not iteratively "improve" the photograph. The first restoration is usually faithful. The third one starts to feel like the model's idea of the person rather than the person.
2. Recreating from a tiny or low-quality file
This is where it gets delicate. If your only photograph of someone is a 320×240 JPEG, an identity-preserving model can plausibly generate a 2048×2048 version that "looks like" them. But the model is inventing pixels that were never recorded by a camera, and those invented pixels are guesses about what their skin, their hair, their eye color must have been.
The rule I use: the smallest source file I will work from is one where I can clearly see both eyes, the corner of the mouth, and the line of the nose. If the source is so low-resolution that the model is guessing the shape of the eyes, you are not restoring the person; you are letting the model decide who they were. That is the line.
Color is the other place to be careful. If the source is heavily yellowed or cyan-shifted from age, color-correct it manually first based on a reference photograph from the same era. Do not let the model "white-balance" a 1978 print by itself. It will guess wrong about skin and you will not be able to unsee it.
3. Generating a new photograph
This is what most people actually want when they say "AI restoration of someone who has passed." Not technically restoration — generation. A new image of the person in a scene that was never photographed: at the wedding they did not live to see, holding the grandchild born after, on the porch of the house they never visited.
A few principles, hard-won:
- Choose a moment, not a fantasy. "Dad on a Saturday morning with coffee" produces something honest. "Dad meeting his great-grandchildren in heaven" produces something the model will render in a way that feels off, because the model has no honest reference for the second prompt.
- Use the best reference photograph you have, not the most flattering one. The model will preserve what it sees. If the reference shows the person tense or unwell, the output will carry that. Choose a reference where the person looks like themselves on a good day.
- Generate at the time of day they actually liked. My father was a morning person. The version of him I kept is in morning light. Forcing him into golden-hour evening light would have produced a version that was technically beautiful and emotionally wrong.
- Their clothes should be their clothes. Describe the kind of shirt they actually wore. A grandfather rendered in a generic linen suit is somebody else. A grandfather in the soft flannel he wore on weekends is him.
The model I trust most for this in 2026 is the InstantID + Flux-Kontext combination. It preserves face geometry strongly even at low reference resolution, and it does not over-smooth skin texture the way some earlier models did. (Smoothed skin reads as "AI" to families almost immediately — pores and small lines are what makes a face read as real.)
What to do with the photograph after
Print one copy. Write the date on the back. Keep the source files together with the original reference photograph. Do not post it publicly without thinking about who might see it — a sibling's grief is theirs, and a public Facebook post may not be the way they wanted to encounter their parent's face again.
I keep my father's portrait in a small frame on the bookshelf next to the 2018 wedding photo. The frame is the same. The wedding photo is still the one most people notice first when they come into the room, because it is the one they remember from the funeral program. The new one is for me.
If you want to try this
DuoPortrait has a Remembrance mode designed specifically for this use case. The interface is quieter than the standard flow — no emojis, no exclamation marks, slower copy. You upload one reference photograph of the person, choose a scene, and receive a single image at print resolution. The first photograph is free. Source files are deleted after seven days. The photograph is never used to train any model.
I built this mode because the standard "AI photoshoot" tone was wrong for what people were actually using these tools for. Grief deserves a slower interface.
If you want to read more about grief and the role of images in it, the Modern Loss essay collection is the best writing I know of on the subject. It is what I read in 2022.
Take care of yourself. If you make a picture, make one.
Mia Tanaka is a portrait photographer and visual-AI designer. She founded DuoPortrait after her own family experience with photographs and loss.
