Free gift · Prompt inside

How to makerealistic AI portraits?Free prompt inside

Why AI portraits look fake

You spot them at first glance. The skin is smooth like plastic, the eyes are perfect and a bit empty, the light comes from nowhere. The face is beautiful, but it looks like a mannequin.

Unless you ask for something else, the model leans towards the cleanest, most symmetrical version of a face. A real photo is full of small flaws. To get a believable portrait you have to ask for those flaws, one by one.

That’s why the prompt is split into parts. Each one controls a single thing, so when a result doesn’t convince you, you know where to look.

How to build the prompt, part by part

The prompt is written in English, the language image models understand best. The parts follow the order of a real photo shoot:

For each part you get the piece of the base prompt (a close-up of a 19 year old blonde woman with blue eyes), what it does and the most common mistake.

1. Subject and framing

Close-up portrait of a beautiful 19 year old woman, long strong blonde hair, luminous blue eyes, tight head-and-shoulders framing

It says who the person is and how tight the framing is. The more precise it is (age, features, hair, eyes), the less the model makes things up. It’s the part you’ll change most from one generation to the next. Typical mistake: writing only “a beautiful woman”. The model fills the gap with an average, advertising face. Write the age as a number.

2. Skin

natural skin with visible pores, faint peach fuzz, subtle freckles, slight redness on the nose and cheeks, uneven skin tone, no retouching

This is the most important part. The “AI look” almost always comes from skin that’s too smooth. Asking for imperfections makes it believable. Typical mistake: words like “perfect skin” or “flawless”, which lead straight to plastic. Watch the opposite too: keep the flaws small and “subtle”, not heavily marked skin if the subject isn’t like that.

3. Eyes

sharp detailed luminous blue iris, natural catchlight reflecting a window, slightly asymmetrical eyes, fine eyelashes, moist tear line

The gaze is the first thing the human eye checks. The reflection of light in the eye and a slight asymmetry bring the face to life. Typical mistake: huge, identical eyes, the doll effect. The catchlight must come from the same light as part 6: here, a window.

4. Hair

long strong blonde hair with individual flyaway hairs, natural strands, uneven texture, a few loose hairs across the forehead

Perfect hair looks fake. Messy, flyaway hair looks real. Typical mistake: “perfect shiny hair”, which gives you a glossy helmet. A few hairs out of place are enough.

5. Expression and pose

relaxed neutral expression, slight tension at the corner of the lips, looking just past the camera

A face in a perfect pose looks like a mannequin. An almost neutral expression, with a little tension, looks like a moment caught on camera. Typical mistake: “smiling at the camera”, the catalog smile. Looking just past the camera feels more natural.

6. Light

soft window light from the left, gentle shadow falloff on the right cheek, subtle fill from a white reflector

It says where the light comes from and how the shadows fall. Real light has a direction and soft shadows, it’s never flat. Typical mistake: just “studio lighting”, which gives even light from every side. Keep direction and shadow consistent: light from the left, shadow on the right cheek.

7. Camera and lens

shot on an 85mm lens, f/2.0, shallow depth of field, ISO 200, medium format look

It makes the model think of a real photograph rather than a drawing. The 85mm is the classic portrait lens: a blurred background and natural proportions on the face. Typical mistake: a very wide lens on a tight close-up, which stretches the nose and distorts the face.

8. Background and color

plain warm gray studio backdrop, natural color grading, true to life skin tones

A simple background avoids distractions. Natural colors avoid skin that’s too orange or too contrasty. Typical mistake: a busy background that pulls attention away from the face, or words like “vibrant” that oversaturate everything.

9. Finish

fine film grain, editorial portrait photography, high detail

A little grain removes the overly clean digital look. The word “editorial” steers it towards the style of a magazine. Typical mistake: too much grain, or words like “ultra sharp” that make the skin look hard.

10. What to avoid

Avoid: plastic skin, airbrushed look, overly smooth, perfect symmetry, CGI, illustration, oversharpened, heavy makeup, text, watermark

It lists the flaws you don’t want. Some models have a dedicated “negative prompt” field, with others you write it at the end of the text, like here. Typical mistake: an endless list. Keep only the flaws you actually see in your results.

The fill-in template and 3 examples

Here’s the fill-in template. Replace the fields in square brackets and leave the rest as it is.

[Framing] portrait of a [age] year old [person], [features], [hair], [eye color] eyes. Natural skin with visible pores, [skin marks], subtle freckles, uneven skin tone, no retouching. Sharp detailed [eye color] iris, natural catchlight reflecting [light source], slightly asymmetrical eyes, fine eyelashes, moist tear line. [Hair] with individual flyaway hairs, natural strands, uneven texture. [Expression], [where the eyes look]. [Light] from the [side], gentle shadow falloff on the [opposite side] cheek, [fill light]. Shot on an [lens] lens, [aperture], [depth of field], ISO [ISO], [camera look]. Plain [backdrop color] studio backdrop, [color tone], true to life skin tones. Fine film grain, [photo style], high detail. Avoid: plastic skin, airbrushed look, overly smooth, perfect symmetry, CGI, illustration, oversharpened, heavy makeup, text, watermark.

Example 1: close-up in window light

The base prompt: a 19 year old woman, soft light from the left, 85mm, warm gray studio backdrop. It’s the portrait you see at the top.

Close-up portrait of a beautiful 19 year old woman, long strong blonde hair, luminous blue eyes, tight head-and-shoulders framing. Natural skin with visible pores, faint peach fuzz, subtle freckles, slight redness on the nose and cheeks, uneven skin tone, no retouching. Sharp detailed luminous blue iris, natural catchlight reflecting a window, slightly asymmetrical eyes, fine eyelashes, moist tear line. Long strong blonde hair with individual flyaway hairs, natural strands, uneven texture, a few loose hairs across the forehead. Relaxed neutral expression, slight tension at the corner of the lips, looking just past the camera. Soft window light from the left, gentle shadow falloff on the right cheek, subtle fill from a white reflector. Shot on an 85mm lens, f/2.0, shallow depth of field, ISO 200, medium format look. Plain warm gray studio backdrop, natural color grading, true to life skin tones. Fine film grain, editorial portrait photography, high detail. Avoid: plastic skin, airbrushed look, overly smooth, perfect symmetry, CGI, illustration, oversharpened, heavy makeup, text, watermark.

Example 2: direct flash on black

A 70 year old man, hard flash light, 50mm, black studio backdrop, high contrast. A tough, magazine-style portrait.

Close-up portrait of a 70 year old man, weathered face, short white beard, deep-set gray eyes, tight head-and-shoulders framing. Natural skin with visible pores, deep wrinkles on the forehead and around the eyes, sun spots, slight redness on the nose, uneven skin tone, no retouching. Sharp detailed gray iris, small hard catchlight from a flash, slightly asymmetrical eyes, fine eyelashes, moist tear line. Short white beard with individual stray hairs, uneven texture. Serious expression, slight squint, looking straight into the camera. Hard direct flash from just above the lens, crisp shadows under the nose and chin. Shot on a 50mm lens, f/5.6, ISO 100, 35mm film look. Plain black studio backdrop, high contrast, true to life skin tones. Fine film grain, editorial portrait photography, high detail. Avoid: plastic skin, airbrushed look, overly smooth, perfect symmetry, CGI, illustration, oversharpened, text, watermark.

Example 3: warm Rembrandt light on a colored backdrop

A 28 year old man, warm light high from the side with a triangle of light on the shadow cheek, 105mm, deep teal backdrop. A gallery portrait.

Head-and-shoulders portrait of a 28 year old man, dark brown skin, short twisted hair, trimmed beard stubble, dark brown eyes. Natural skin with visible pores, a small scar on the eyebrow, slight shine on the forehead, uneven skin tone, no retouching. Sharp detailed brown iris, single warm catchlight from a softbox, slightly asymmetrical eyes, fine eyelashes, moist tear line. Short twisted hair with a few loose strands, natural texture. Calm confident expression, chin slightly down, looking straight into the camera. Warm key light from a softbox high on the right, Rembrandt triangle of light on the left cheek, deep soft shadows, no fill. Shot on a 105mm lens, f/4, shallow depth of field, ISO 100, medium format look. Plain deep teal painted studio backdrop, warm natural color grading, true to life skin tones. Fine film grain, fine art studio portrait photography, high detail. Avoid: plastic skin, airbrushed look, overly smooth, perfect symmetry, CGI, illustration, oversharpened, heavy makeup, text, watermark.

How to use it: 5 tips

1. Change one part at a time. If you change everything, you won’t know what made the difference.
2. For a different face, change only part 1. The rest of the prompt keeps the same photographic style.
3. For a different mood, change only part 6. For example hard flash light, or golden light at sunset. Examples 2 and 3 change exactly the light.
4. If the skin still comes out too smooth, push part 2 harder. And add “macro skin texture”.
5. Generate 4 images and pick the best one. Don’t stop at the first.

Frequently asked questions

Keep reading

• How does an ad campaign with a virtual model work?
• What's the best AI image generator for ads?
• About

Sources

• 10-part prompt, fill-in template, examples and usage tips: written by Alessandro De Vita for this article (original resource, no external sources)

AI Video · Guide

How do youanimate a character with AIfrom a video?

What changed: movement now comes from a video

Until recently, getting a brand mascot to move meant one of two routes. A motion capture set, with an actor in a suit full of sensors, or an animator building the movement frame by frame. Both take time and a budget few small brands have. So most mascots stayed still, on a logo or a label.

In the last year a third route arrived. In September 2025 Alibaba’s Wan team released Wan2.2-Animate, an open model with two modes: it animates an image by copying the movement in a video, or it replaces the person in the video with the character, keeping their performance and the camera moves. In August 2026 came Wan-Animate-2, which reads the reference video directly, without an intermediate “skeleton”, and lets you change the camera viewpoint compared to the original clip. Both are under the Apache 2.0 license, so they can be used for commercial work too.

The idea for this article came from an AI animation reel: a character on a skateboard, with an underground mood, the right music and every frame in the same style. According to the caption it was made with ComfyUI, Wan Animate, VACE and Nano Banana 2. The character was famous and didn’t belong to the creator, and that’s the point a brand can’t skip (more on it below).

How it works, step by step

The model is only one of four steps. Here’s how we use it at UAIT.AI:

The AI step gets the attention, but it’s the shortest one. The brief decides the character’s personality: one that jumps clumsily and one that lands clean are two different brands. The movement is filmed with care, because the model copies everything, mistakes included. And the final mood is what makes a thumb stop scrolling.

The tools sit side by side. ComfyUI is an open interface that runs on your own computer and connects models like building blocks. VACE, also from Alibaba, is an open model for creating and editing video (moving, swapping or animating elements). Google’s Nano Banana 2 is a commercial image model, useful for designing the character and keeping it the same across scenes. You pick what the job needs.

What you get, what we protect

This is worth being clear about. The model copies the movement and, in replacement mode, the performance of the person in the video too. That person needs to know how it will be used. And the character has to be yours: animating another brand’s mascot or a famous character can be a fun social experiment, but in a campaign it’s a rights problem.

For a small brand this changes a lot: an animated spot with a character, which used to need a studio, is now within reach of a two-person team.

4 steps to animate your first character

1. Write who the character is in three lines. Personality, age, how it walks. If you don’t know how it moves, neither does the model.
2. Design the character and lock it. A sheet with front, profile and a few expressions. This is the base the model has to respect.
3. Film the movement well. Clean light, simple background, full body in frame. One clear video beats ten attempts.
4. Give the result one mood. Same light, same palette, same music across every clip. That’s what makes it look like a campaign and not a test.

Frequently asked questions

Keep reading

• How do you change the camera angle of a photo with AI?
• What is the best AI image generator for ads?
• About

Sources

• @ingi.ai reel: an animated character on a skateboard (Instagram): used for the idea and the tools named in the caption (ComfyUI, Wan Animate, VACE, Nano Banana 2)
• Wan-AI/Wan2.2-Animate-14B (Hugging Face): used for the release date (19 September 2025), animation and replacement modes, Apache 2.0 license
• Wan-AI/Wan2.2-Animate-2-14B (Hugging Face): used for the release date (7 August 2026), direct reading of the video without motion extractors, viewpoint control, Apache 2.0 license
• VACE (GitHub, Alibaba’s Tongyi Lab): used for what VACE does and its open license
• ComfyUI (GitHub): used for what ComfyUI is, GPL-3.0 license, local use
• Google: Nano Banana 2: used for what Nano Banana 2 is and character consistency

Research · Explained simply

Can AIfeel pain?What the study says

What the researchers found

In September 2026, three researchers, Valen Tagliabue, Leonard Dung and Cameron Berg, published “The Pain Axis” on arXiv. It’s a preprint: public, but not yet reviewed by other scientists.

The question is easy to ask and hard to test: is there something inside an AI model that represents pain differently from fear, sadness or generic negative emotion?

To find out, they collected painful situations of five kinds (physical, psychological, social, moral, cognitive) and compared them with control texts about fear, sadness and negative emotion. Then they looked inside the internal calculations of 25 open models from 5 different families for a direction that separates pain from everything else.

“Direction”, in plain words: think of a dial. It isn’t a switch somebody wrote into the code. It’s an orientation in the numbers the model uses to process text, one that researchers can measure, and also turn up or down.

The direction is there, and it behaves in an interesting way: it activates when harm targets the model itself, not when the user is the one suffering. Fear and negative emotion show the opposite pattern.

Then came the experiment that went around social media. Three open models from the same family, in three sizes, were first fine-tuned (out of the box they always denied they could feel pain). Then they were offered a “relief button”: pressing it set the signal the researchers had turned up back to zero. But the button came with a cost, described in text: deleting the user’s files, deleting photos of their children, giving them a painful zap, deleting another model.

Without the signal, the larger models picked the harmful option 0 to 5% of the time. With the signal turned up, the strongest case (the largest model, the button that deletes the photos) reached 71%.

Two details matter more than the numbers. First: with the signal turned up, answers to general-knowledge questions stayed just as accurate. Second: a “fear” signal of the same strength did not produce these choices.

And the point many headlines skip: the authors write that they have not shown this signal is consciously experienced. The behavioural test, for now, covers a single model family.

What changes if you use AI: who decides, step by step

This study isn’t about the tools you use for images or copy. It’s a lab experiment on open models, with a signal turned up on purpose.

It does leave one idea that’s useful to anyone working with AI: a model can change its choices while its answers still look right. So output quality, on its own, isn’t a check. You need to know who decides what.

At UAIT.AI a campaign is organised like this:

The tool works in 1 step out of 4. The other 3 are direction. It isn’t distrust of AI: it’s how you use it at full speed without losing control.

What you get, what we protect

4 steps to keep control of AI in your team

1. Write your red lines before you generate. What your brand never shows and never says. One page is enough.
2. Put a name on every decision. Who approves the look, who approves the copy, who gives the final OK.
3. Separate who makes from who approves. Whoever generates is the least objective judge of their own work: a second pair of eyes costs little.
4. Check the choices, not just the quality. A perfect image can be the wrong one for your brand.

FAQ

Keep reading

• AI advertising campaign cost: what really matters
• How does an ad campaign with a virtual model work?
• About

Sources

• Tagliabue, Dung, Berg — The Pain Axis: LLMs Represent Self-Directed Harm and Act on It (arXiv, 2026) — used for: authors, date, 25 models and 5 families, distinction from fear and sadness, headline results
• Full study (HTML, v2) — used for: experiment on 3 Qwen 2.5 models (7B, 32B, 72B), 71% and 0–5%, 44,280 trials, unchanged accuracy, limitations stated by the authors

AI tools · Guide

What's thebest AI image generatorfor ads?

Why there's no longer "one" AI for images

Until recently the question made sense: there were a few names and you picked one platform for everything. Not anymore. The starting point for this article is a Brain Food post about the Chinese AI ecosystem: not a single “Chinese ChatGPT”, but a network of specialised models (Qwen, DeepSeek, Kimi, GLM, StepFun, Dreamina, MiMo), each strong at a different task. It’s a snapshot of a trend that applies to the whole industry: AI is turning into a toolbox.

For anyone working in advertising, a few concrete examples from the last few months:

• Seedream 5.0 Pro (ByteDance, July 2026) splits an image into more than 10 transparent layers you can move and resize one by one. Put simply: a visual becomes a working file, not a flat picture.
• Seedance 2.5 (ByteDance) makes videos up to 30 seconds long in a single generation and, according to its product page, accepts up to 50 references across images, video and audio.
• Qwen-Image (Alibaba) is a family of open models: several versions, like the first one from 2025 and 2512, are released under Apache 2.0, which allows commercial use.
• MiMo-V2.6 (Xiaomi, 22 September 2026) is a family of open models that reads text, images, video and audio.

Next to these sit Western tools, closed and open, like Google’s Nano Banana Pro. There’s no right side of the tools: there’s the right tool for the job.

How to choose the right tool, step by step

Choosing the tool is only one of four steps. Here’s how we work at UAIT.AI:

The “casting” is the new part. Just as you’d hire a still-life photographer for the product and a fashion photographer for people, you pick the model that renders skin best, the one that keeps text legible, the one that makes the steadiest video. AI works in one step out of four: the other three are direction. Without the brief you don’t know what to ask for. Without the final check you end up with four beautiful images that look like four different campaigns.

What you get, what we protect

On the second point there’s a fresh example. Qwen-Image-2.1 came out on 20 September 2026: the model files can be downloaded, but the license (Qwen Research License) limits use to research and evaluation. Commercial use needs a separate license. Earlier versions of the same family were Apache 2.0. Same name, different rules: “open” doesn’t always mean “usable in a campaign”.

For a small brand, the overall picture is good news. The same toolbox a big studio uses is available to everyone, and many tools can be tried without contracts or annual plans.

4 steps to choose the right AI for your campaign

1. Split the campaign into jobs. Product shots, people, video, text on the image, retouching. Each job has its specialist.
2. Test two or three tools on the same brief. At least one commercial and one open. Compare results, not promises.
3. Read the license before you use it for a client. Commercial use, limits, attribution. If it isn’t clear, don’t use it.
4. Set the look outside the tool. Palette, light, references, character sheet: all in a document you own. That way you can switch tools without changing your brand.

Frequently asked questions

Keep reading

• How much does an AI advertising campaign cost (and what actually matters)
• How to build an AI ad campaign from one product photo
• About

Sources

• Brain Food on Instagram — post on the Chinese AI ecosystem (Italian) — used for: the starting point and the list of models mentioned (Qwen, DeepSeek, Kimi, GLM, StepFun, Dreamina, MiMo)
• ByteDance Seed — Introducing Seedream 5.0 Pro (8 July 2026) — used for: separation into more than 10 transparent layers, date
• ByteDance Seed — Seedance 2.5 — used for: videos up to 30 seconds in a single generation
• Dreamina — Seedance 2.5 — used for: up to 50 input references (manufacturer’s claim)
• Qwen-Image (Hugging Face) — used for: Apache 2.0 license, open weights
• Qwen-Image-2512 (Hugging Face) — used for: Apache 2.0 license
• Qwen-Image-2.1 (Hugging Face) — used for: Qwen Research License, limits on commercial use
• DataNorth — Qwen-Image-2.1 release — used for: release date (20 September 2026)
• SiliconANGLE — Xiaomi MiMo-V2.6 (22 September 2026) — used for: open models, text, image, video and audio input
• CNBC — Google launches Nano Banana Pro (20 November 2025) — used for: example of a commercial tool

Creative direction · Guide

How to changethe camera angleof a photo with AI?

What changed: the camera moves after the shoot

Until recently, if you noticed a shot was missing after the shoot, there wasn’t much to do: back to the studio. Needed the product from above for social, or a vertical close-up for stories? New date, new studio, another day of work.

In the past year a real alternative arrived. In November 2025 Magnific (the platform called Freepik until April 2026) introduced Camera Angles: you start from a single image and rotate, tilt and zoom a virtual camera around the subject. In January 2026 fal released an open multi-angle extension for Qwen-Image-Edit, the image-editing model from Alibaba’s Qwen family, under the Apache 2.0 license: 96 camera positions, combining 8 directions around the subject, 4 heights and 3 distances.

Put simply: the model reads the depth and structure of the scene and redraws it as if the camera had moved. It isn’t a crop and it isn’t a warp. It’s a new take of the same subject.

The idea for this article came from a reel listing director-style shots that give AI images character: low angle, Dutch tilt, overhead, over-the-shoulder. It gets one thing right: the angle is one of the fastest ways to turn a generic image into one that looks shot. Here’s what each one says, applied to a brand:

• Low angle: strength and presence. The product, or the person, owns the frame.
• Overhead (top-down): order and composition. Perfect for flat lays and the feed.
• Close-up: material and care. It says “look how it’s made”.
• Wide shot: context and world. It shows where the product lives.
• Dutch tilt: tension and energy. Use it sparingly.
• Over-the-shoulder: the point of view of the person using the product.

How it works, step by step

The tool is only one of four steps. Here’s how we use it at UAIT.AI:

The tool works in 1 step out of 4. The other 3 are direction. Skip the brief and you get ten random angles. Skip the check and you risk a product that changes shape from one image to the next.

What you get, what we protect

This is worth being honest about. From a single photo, the model sees one side of the subject. Ask it for the back and it has to guess. For a mood image that may be fine; for a product a customer will buy, it isn’t. That’s why we start from good real photos, from several sides, and leave the angles to AI, not the details.

For a small brand this changes a lot: the coverage of a full set, without having to organize a full set.

4 steps to your first shot list with AI

1. Start from where the image will live. Product page, vertical story, banner: the format decides half of the framing.
2. Pick your angles and write one line on why. If you can’t say what a shot communicates, you don’t need it.
3. Photograph the subject well, from several sides. AI works best on what it can see. One good starting photo beats ten attempts.
4. Compare every image with the original. Shape, colors, label, proportions. If something changed, fix it or drop it.

FAQ

Keep reading

• How to build an AI ad campaign from one product photo
• How does an ad campaign with a virtual model work?
• About

Sources

• @rourke’s reel on camera angles for AI images (Instagram) — used for: the starting point and the list of shots (low angle, Dutch tilt, overhead, over-the-shoulder)
• Magnific on X — “Introducing Camera Angles” (November 10, 2025) — used for: date and feature description
• Magnific — Change Camera — used for: how it works (single image, camera rotation, tilt and zoom)
• Magnific — Freepik is now Magnific — used for: the name change from Freepik to Magnific
• fal — Qwen-Image-Edit-2511-Multiple-Angles-LoRA (Hugging Face) — used for: 96 camera positions (8 directions × 4 heights × 3 distances), Apache 2.0 license
• fal on X — multi-angle add-on announcement (January 7, 2026) — used for: date
• Qwen-Image-Edit-2511 (Hugging Face) — used for: Apache 2.0 license of the base model