How do you make AI-generated images look like vintage family photographs?

AI images read as vintage family photographs when you do three things: prompt the era specifically — decade, named film stock ("faded film look, think Kodak Portra 800"), candid framing, period clothing — add negative constraints against hyper-sharp modern AI detail, then degrade the output in post with film grain, film damage, a dust overlay around 8% opacity, and a faded, warm-shifted grade.

Start with era-specific prompt language, not the word "vintage." Describe the decade, the film stock, and the occasion — "candid 1970s family gathering, faded film look, think Kodak Portra 800, soft focus, slight overexposure" — because generic style keywords fail the same way "make it cinematic" fails: you only get the backlit warmth and muted contrast you actually describe. Add period-correct clothing, props, and composition cues — off-center framing, subjects mid-motion or mid-laugh, cluttered domestic backgrounds — since candidness is what separates a family photograph from a posed portrait.

invideo is an agentic video creation tool with the current image models available, and the documented workflow there runs in two passes: instruct the invideo agent to generate era-appropriate candid photographs first, then instruct it to degrade the image quality to match the period — fading, soft edges, colour cast — so the set reads as one consistent album rather than clean renders with a filter on top.

If you have real family photos, use them as references the right way: don't ask the model to replicate them. Have the invideo agent read their colour palette and texture qualities and translate those into prompt language — in one documented production this approach returned generations "hyper-realistic with the exact colour temperature" the director wanted, where dropping the reference images directly into prompts did not work.

Model choice matters most for faces. Recraft renders skin imperfections — pores, lines, stubble — that keep AI faces from looking airbrushed, GPT-Image-2 is the pick when the image carries text (a handwritten date, a storefront sign) because it doesn't distort lettering, and Nano Banana handles compositing multiple family members into one frame. All of these run inside invideo, so the invideo agent routes each image to the right model rather than you picking a tool per shot.

Constrain the modern look out with negative prompts: prohibit hyper-sharp detail, digital colour, studio lighting, and modern objects. Tailor the constraint list per image rather than reusing one universal list, and apply negatives sparingly — over-loading them can make the model do exactly what it was told not to do.

Finish with physical degradation in post. AI generations come back ultra-sharp with a plasticky skin quality, and the documented fix is a small blur, grain, and a grade pushed toward the period. In DaVinci Resolve: apply the Film Grain effect at high strength and opacity with a small grain size, add Film Damage for moving lines and specks, layer a dust overlay at around 8% opacity, then warm or sepia-shift the grade and lift the blacks so nothing sits at pure contrast.

Generate in batches and select, since image generation is cheap — one documented workflow requests three grid options per round instead of single images — and keep only the frames that feel accidentally shot. One production built a full collection of vintage family photographs exactly this way for a projector sequence, then animated the stills with four parallel Seedance 2.0 animation sub-agents using 8mm home-movie style instructions — the same photo set can carry into motion if your film needs it.

Watch some of these to see what works for you:

What we tend to do is put a tiny bit of blur on top of the scene, add a bunch of grain and then play with the grade till it comes closer to live action film.

— invideo's creative team