Photoshop Face Swap in 2026: I Put a New Face Into a 400-Year-Old Painting

You know the drill. Lasso the face. Paste it on a new layer. Free Transform, Warp, nudge, nudge again. Load the selection, contract five pixels, Auto-Blend, then fight Curves for twenty minutes because the skin went red, then grey, then red again.

Two hours later you zoom to 100% and there it is: the hard edge along the jaw, the mismatched lighting, the dead stare. A sticker on a photo. Everybody can tell.

Here is the thing nobody says out loud. In 2026, the pros are not doing it that way anymore. The hard 90% (skin tone, lighting, a believable seam) gets handed to an AI model in one step. Photoshop only does the fun 10% at the end.

To prove it, I picked the most unforgiving base image I could think of: the most famous face in art history. Here is how to run a proper photoshop face swap on Girl with a Pearl Earring, keep the oil-paint look, and then make her blink.

Key takeaways

• The classic manual method is roughly 10 steps and 30 minutes to 2 hours per face, and skin tone plus lighting are where it falls apart.
• Photoshop's own Generative Fill and Firefly cannot swap in a specific person's face from a photo you provide. That is by design.
• The 2026 workflow is a hybrid: one AI model fuses the new face in seconds, then Photoshop does the polish.
• Fusion keeps the source material's texture. A good result still looks like an oil painting, not a clean HD photo pasted on canvas.
• The finished still can become a living-portrait video and a shareable GIF without touching a timeline.

A gallery lit warmly at night: a classical oil portrait on the wall whose painted face is half-dissolving into a modern human face, a laptop open on a side table showing a face-swap interface

See It First: A Photoshop Face Swap That Started as a 400-Year-Old Painting

Before any theory, here is the payoff. Left is Johannes Vermeer's original Girl with a Pearl Earring (1665, public domain). Right is the same painting after the face swap. New person, same brushstrokes, same soft light from the left, same craquelure in the varnish. Before and after: on the left the original Vermeer painting, on the right the same painting with a different person's face fused in while keeping the oil-paint texture

That face on the right is not the girl in the original painting. It is a completely different person. But it does not read as a paste job, because the face was re-rendered inside the painting's style instead of dropped on top of it. That difference, fusion versus paste, is the whole game. Keep reading and you can reproduce this exact result.

Why Face Swap in Photoshop Is Still So Hard

The manual method is not a mystery. The canonical version, taught by Jesús Ramirez, goes like this: Lasso the source face, paste it over the target, Free Transform and Warp to line up the eyes and mouth, add a layer mask, load the body layer as a selection, contract it by about 5 pixels, then run Edit ▸ Auto-Blend Layers with Seamless Tones and Colors checked, and finally clean up with Curves, Selective Color and a bit of dodge and burn (Photoshop Training Channel, 2024).

It works. It also breaks in the same three places every time.

• Skin tone. Match Color and Curves get you close, then the face reads slightly red, or slightly grey, and no slider quite fixes it. One long-running Adobe forum thread on the topic is literally titled "nothing works as advertised."
• Lighting. If the light in the source face comes from a different direction than the target, the brain flags it instantly. People describe the output as looking "like a mask, not a face."
• The eyes. A dead, glassy stare is the fastest way to land in the uncanny valley, and it takes real dodge-and-burn skill to fix.

There is also a limit people run into and assume they are doing something wrong. Photoshop's Generative Fill and Firefly can regenerate a face, but they cannot swap in a specific person's face from a photo you hand them. Firefly only generates from Adobe's own trained material, so "replace this face with my friend's face from this other picture" is not something it will do (Adobe Community, 2023). That single limitation is why so many people go looking for a different model. An obviously bad manual face swap example, with a mismatched skin-tone patch, a hard cut-out edge along the jaw, and a flat lifeless stare, marked with red callouts

A Faster Photoshop Face Swap Workflow (AI + Photoshop)

Here is the shift. Instead of doing the fusion by hand, you let a modern image-editing model do it, then you finish in Photoshop.

This is not a fringe idea. As of 2025, Photoshop's own Generative Fill can already call third-party models like Google's Nano Banana and Flux instead of Firefly, because Adobe knows those models are better at certain jobs (Digital Camera World, 2025). The catch is that the in-app version is throttled and still will not take your reference face. So the practical move is to run the same class of model directly, at full resolution, with your own two images, in one browser tab.

That is where Atlas Cloud comes in. It is a model playground where the exact models Photoshop reaches for (Nano Banana 2 for the edit, plus image-to-video models) run without the app's limits, and you can feed it your own reference photo. The whole chain below happens in one tab: fuse the face, then animate the result. A two-by-two workflow grid: your face photo, the original Vermeer painting, the fused still, and one frame from the finished living-portrait video

Two jobs, two models. Here is what each one does and what it costs.

How to Face Swap in Photoshop, Step by Step

This is the full chain, start to finish, using one showcase image. Every prompt below is copy-paste ready.

Step 1: Prep your two images in Photoshop. Download the base you want to swap into. For this demo it is the public-domain Vermeer from Wikimedia Commons. Then pick your face photo. A front-ish angle with soft, even light works best, ideally lit from roughly the same side as the painting (here, the left). In Photoshop, crop out clutter, straighten both images, and get them to a similar resolution. No AI yet, this is just housekeeping so the fusion has clean input. A Photoshop workspace with the Vermeer painting and a separate portrait photo open side by side, lasso and crop tools visible (illustrative)

Step 2: Do the hard part with one AI fusion. Open the google/nano-banana-2/edit model, load two reference images (Image 1 = the painting, Image 2 = your face), set the resolution to 2K, and match the output ratio to the painting. Paste this exact prompt:

plaintextCopy
1Image 1 is a classic 17th-century Dutch oil painting: a young figure against a
2dark background, wearing a blue-and-gold headscarf and a large pearl earring,
3lit softly from the left, rendered in visible oil brushstrokes with fine
4craquelure. Image 2 is a photo of a person's face. Perform a complete face
5swap: replace the painting's face with the face and identity of the person in
6Image 2. This identity change is required, do not keep the original face.
7Re-render the new face fully in the same oil-painting style: same soft left
8lighting, same warm skin palette, same visible brushstrokes, canvas texture and
9craquelure, the same slight head-turn and direct gaze. Keep the headscarf, pearl
10earring, dark background, framing and painterly grain exactly unchanged. The face
11must stay realistically human.
12

Settings: resolution 2K, two reference images, ratio matched to the painting. Nano Banana can vary between runs, so if the first output loses the brushstrokes or drifts off the pose, run it again. You are done when the swapped face is clearly a different person but still looks hand-painted. Nano Banana 2 Edit on Atlas Cloud: both reference images loaded on the left, the fused oil-painting result rendered on the right

Step 3: Finish in Photoshop (the pro 10%). Bring the fused result back into Photoshop. Now you do only the polish: nudge the last bit of skin tone with Curves or Selective Color, unify the light with dodge and burn, heal any faint seam with the Healing Brush, add a touch of noise/grain so it matches the canvas texture, and sharpen the eyes just enough to kill any glassiness. If you stacked layers, Edit ▸ Auto-Blend Layers with Seamless Tones and Colors is a good final pass. This is the part manual editors spend two hours on. Here it takes a few minutes because the fusion already did the heavy lifting. A Photoshop adjustment-layers panel with Curves and Selective Color open over the fused portrait, and the Auto-Blend Layers dialog with Seamless Tones and Colors checked (illustrative)

Step 4: Bring the portrait to life. Take the finished still and open bytedance/seedance-2.0-mini/image-to-video. Set the image to your fused portrait, pick 720p, and a 5-second duration. This is the 2026 living-portrait trend that museums and photographers have picked up. Paste this prompt:

plaintextCopy
1A living portrait. The subject in this oil painting slowly comes to life: one
2soft, natural blink, the faintest breath, an almost imperceptible turn of the
3head toward the viewer, eyes catching the light. Preserve the oil-painting
4texture, brushstrokes and warm gallery lighting; locked-off camera, still
5background, cinematic and subtle, no exaggerated movement.
6

Settings: 720p, 5 seconds, ratio matched to the painting. Keep the motion subtle. The magic is that it still looks like a painting that happens to breathe, not a cartoon. Seedance 2.0 Mini image-to-video on Atlas Cloud: the fused portrait loaded as the input frame, the finished 5-second clip playing in the output panel

Step 5: Turn the video into a shareable GIF. Atlas Cloud hands you an MP4, not a GIF. To get a clean, looping GIF for social, pass the video to any LLM that can run code (Claude, ChatGPT, or Atlas Cloud's LLM API) with this prompt:

plaintextCopy
1Here is a 5-second MP4 (attached / at <URL>). Convert it into a high-quality,
2infinitely-looping GIF under 5 MB for web and social. Use ffmpeg with a two-pass
3palettegen/paletteuse: first generate a palette, then apply it at ~480px width,
412 fps, with dithering, so the colors stay clean and the oil-painting texture is
5preserved. Return the exact ffmpeg commands and the final .gif.
6

If you would rather run it yourself, the two-pass ffmpeg recipe is short:

plaintextCopy
1ffmpeg -i in.mp4 -vf "fps=12,scale=480:-1,palettegen=max_colors=128" palette.png
2ffmpeg -i in.mp4 -i palette.png -lavfi "fps=12,scale=480:-1,paletteuse=dither=bayer" -loop 0 out.gif
3

And here is the final result: a 400-year-old painting, a brand new face, blinking and breathing on a loop. The finished living-portrait GIF: the swapped-face Vermeer painting blinking and breathing subtly on a loop

INRmWve2HVk

More Photoshop Face Swap Variations to Try

The painting was the hard mode. The same workflow handles cleaner cases with even less effort.

• Photo into a real portrait. Want to put your own face into a crisp photograph (say, an official-looking studio portrait) instead of a painting? For that, atlascloud/face-swap-image is the better tool. It preserves the original composition and transfers a real identity, rather than re-rendering the whole scene.
• Magazine-cover studio look. Start with a clean studio portrait as the base and the result reads like a real cover shoot.
• Batch or multi-face. Swapping several faces in one frame? Run each face as its own pass rather than asking one edit to juggle two people, then composite.

Which model for which job:

What a Photoshop Face Swap Actually Costs

The honest comparison is not "free versus paid," it is "your afternoon versus a few cents."

You still need Photoshop for the polish and for full creative control. But the part that used to eat your afternoon now costs less than a coffee and finishes before you have refilled it.

Is Face Swapping Legal? A Quick Note

Short version: swap your own face, public-domain art, or a face you have permission to use, and you are fine. That is exactly why this tutorial used a 1665 painting and a generated portrait. Putting a real person's face somewhere without consent is a different matter, touching on likeness rights and a growing set of deepfake laws. It is not a hypothetical problem either. Deepfakes made up 7% of all fraud attempts in 2024, a four-fold jump from the year before (Sumsub, 2024). Use the tech for art, restoration and fun, not for passing something off as real. Atlas Cloud also blocks sensitive swaps, including anything involving minors.

Frequently Asked Questions

How do I swap faces in Photoshop step by step?

The manual route: Lasso the source face and paste it over the target face on a new layer, use Free Transform and Warp to align the eyes and mouth, add a layer mask, load the underlying body layer as a selection and contract it by about 5 pixels, run Edit ▸ Auto-Blend Layers with Seamless Tones and Colors checked, then finish with Curves, Selective Color and dodge and burn. It is roughly ten steps. The faster route is to let an AI model fuse the face first, then only do the polish in Photoshop.

Can Photoshop's Generative Fill or Firefly swap in a specific person's face?

No. Generative Fill and Firefly can generate a new, invented face, but they will not take a specific face from a photo you provide, because Firefly only generates from Adobe's own trained material. This is exactly why people use an edit model that accepts your own reference photo, like running Nano Banana 2 on Atlas Cloud, for a real photoshop face swap.

How do I match skin tone and lighting so the face swap doesn't look fake?

Manually, the tools are Auto-Blend Layers (Seamless Tones and Colors), Curves, Selective Color and dodge and burn to unify the light. The faster answer is to let the fusion model match tone and lighting in one step, then only touch up in Photoshop. The two tell-tale signs of a fake are a hard edge along the jaw and a face lit from the wrong direction, so fix those first.

Is there a faster or free alternative to a manual Photoshop face swap?

Yes. A single AI fusion does in seconds what takes 30 minutes to 2 hours by hand. It is not free, but a fused still costs cents, and you keep Photoshop for the creative polish. For most people the hybrid workflow is faster, more realistic, and cheaper in time than doing the whole thing manually.

Can I turn my Photoshop face swap into a video or GIF?

Yes, and it is the fun part. Feed the finished still into an image-to-video model (like Seedance 2.0 Mini) to get a short living-portrait clip, then convert that MP4 to a looping GIF with a two-pass ffmpeg palettegen. The article above has the exact prompts and the ffmpeg commands.

How do I swap faces in Photoshop while keeping an oil-painting look?

This is where fusion beats pasting. Use an edit model and tell it explicitly to re-render the new face in the source style: same lighting, same brushstrokes, same canvas texture and craquelure. If you paste a photo face onto a painting and leave it there, it will always read as a sticker. The face swap in photoshop only looks right when the new face is rebuilt in paint, which is exactly what the Step 2 prompt forces.