How to Bring a Painting to Life With Yourself in the Frame: A GPT Image 2 and Runway Gen 4 Turbo Workflow

Today I tested a format built around living paintings, where the viewer first sees a static image, and then that same frame gently comes to life in a short video. It sounds simple in theory. In practice it breaks down in two places: the face starts to distort, and the motion gets too noisy and kills the magic.

Below I walk through the workflow that worked for me across four different paintings. I started by placing my own face into the artwork using GPT Image 2, then animated the frames in Runway Gen 4 Turbo. For publishing, I turned this into a Telegram post, an Instagram carousel, and a comment with the prompts.

You can see the finished example in the portfolio: the living paintings case on evskakunova.ru. If an old cached version loads, here's a direct link with the cache reset: view the case in the portfolio.

Why this format grabs attention

Most AI visuals get scrolled past like a pretty picture. A living painting works differently: the viewer first registers the familiar aesthetic of a painting, then notices that the figure in the frame is alive. That's a small visual shift, and it holds attention well in Reels, carousels, and Telegram feeds.

But the format only holds together through restraint. Give the character a sharp head turn, too wide a smile, or a long action, and the face falls apart while the painting's atmosphere disappears with it. What wins here isn't spectacle, it's precision.

Step 1. Choose paintings that can handle animation

Not every painting is equally suited to animation. The best results come from paintings where the subject already has a clear pose, a calm turn of the head, and a readable emotion. That way the AI doesn't have to invent too much motion on its own.

For today's test I used four different works, so I wouldn't repeat the same angle. It matters that the scenes differ in composition, not just in color.

Step 2. Place your own face in, don't just apply a style

The main task at the image stage isn't to render something in the style of the painting. You need to preserve the person: the eyes, the shape of the face, the smile, the overall sense of who this is. Only after that do you transfer that likeness into the clothing, lighting, and composition of the painting.

I did this in GPT Image 2 with two references: the painting itself and a photo of me. Prompts work well when they explicitly say to preserve the face and identity of the person in the reference, and to take the costume, light, composition, and era from the painting.

Key principle: hold the face first, then the style. Swap that priority and the image might look beautiful, but it won't be you anymore.

Step 3. For video, choose exactly one motion

The most common mistake in animating work like this is trying to make the character perform an entire mini-scene. For paintings, that's almost always too much. The face starts to drift, hands break apart, fabric moves independently of the person.

The workflow that worked for me in a 5-second clip came down to this: one gentle gesture, one emotion, one barely noticeable action. For example: a light nod, a calm smile, a slow glance to the side, an almost imperceptible breath. The quieter the gesture, the cleaner the face holds up.

Why I chose Runway Gen 4 Turbo

For this task, control over the face mattered more to me than the lowest price. The cheapest options are exactly where the problems that matter most here tend to show up: distorted features, unwanted head turns, strange mirror reflections, extra facial movement.

Runway Gen 4 Turbo turned out to be more precise for short, gentle motion in this workflow. I still capped every animation at five seconds with a very narrow action, so I wasn't trading quality for extra dynamics.

Which phrasing helps preserve the face

In my animation prompts, I stripped out anything that pushed the model toward a strong head turn or a complex pose change. Instead of big actions, phrases like gentle smile, soft nod, subtle gaze shift, and barely perceptible movement work better.

It also helps to repeat the constraints as negatives: no sharp turns, no strong facial expressions, no change to the composition, no facial distortion, no extra hand animation.

Turning one test into a full content package

For Telegram, I put together a post with three pairs: a static image first, then its animated version. For Instagram, I built a carousel with the video embedded in the cards, and the funnel led further into the Telegram channel.

I prepared the prompts as a separate comment rather than overloading the post itself with them. That works better both for readability and for engagement: people want to see the result first, and only then go grab the mechanics.

Where to see the result

If you'd rather see the finished result first and break down the mechanics after, open the portfolio with this case. It has the final video for the card, along with the image I used in the publication.

Where Syntx fits in

For this format, Syntx is handy as a single working window, especially when you don't want to jump between several services. Through it you can quickly pull references, test models, and avoid spreading yourself across ten different dashboards.

Try Syntx via my link →

What I wouldn't do next time

I wouldn't put paintings with the same angle too close together in one package. When two frames repeat each other in pose and lighting, the carousel looks weaker as a whole. I'd also avoid motions with a strong profile turn or a mirror reflection, if the goal is to keep the face as clean as possible.

This format wins when every piece differs in look and composition, while each individual card stays calm inside itself.

What's next

If you want to recreate this format for yourself, start with three things: a real painting with a strong female figure, a good portrait photo, and a very quiet 5-second motion script. Everything else is just technique from there.

And if you need more than a prompt (a full build of a package like this ready for publication), book a consultation. This is exactly where I work, at the intersection of AI visuals, content, and funnel packaging.