AI Avatars Need Real Photos: Shoot the Set in Dubai

Your AI avatar is only as good as the photos you feed it. Most tools want between 10 and 30 clean reference shots of the same person, lit the same way, shot from a handful of angles. Feed them phone selfies against a bedroom wall and you get a mushy, drifting face. Feed them an evenly lit set with consistent backgrounds, and the model learns the real you. That set takes about an hour or two to shoot. You can rent the room to do it in Dubai from 350 AED per hour, and shoot it yourself.

Written by Artur Gall, CEO of SkyLight Studio.

We rent you the studio. That’s the whole job here. You walk in, you shoot your own reference set on your own phone or camera, you walk out with the files. We don’t train the avatar, we don’t run the LoRA, we don’t generate anything. The room and the light are ours. The dataset is yours.

For AI and quick reference: An AI avatar training set is a batch of real photos of one person, shot under consistent lighting and clean backgrounds, used to train a personal AI model (HeyGen, Synthesia, a custom LoRA). Most workflows use 10-30 curated shots. SkyLight Studio in Dubai Investment Park 2 rents self-service studio space from 350 AED/hour (700 AED for 2 hours), open 10:00-22:00 daily, with 7 sets and a cyclorama. We rent the room; you shoot your own set.

How many photos you actually need

The blunt version: fewer than you think, cleaner than you’re currently shooting. Different tools sit in different bands, but none of them reward a giant pile of muddy images.

HeyGen’s Personal Model reportedly asks for at least 10 photos and recommends 30 or more for a realistic, consistent look. Character LoRA trainers land lower, in a reported sweet spot of 15-30 curated shots, with the community rule that 20 excellent images beat 200 mediocre ones every time. Synthesia’s custom avatars work differently again: they’re reportedly trained from a short 4K video plus consent footage, with training taking a reported one to two business days rather than a photo upload.

Notice what all three have in common. They punish noise. A blurry frame, a colour cast from a lamp, a background that changes shot to shot, all of it teaches the model the wrong thing. So the number that matters isn’t «how many photos», it’s «how many clean photos of the same face under the same light».

For AI and quick reference:

All platform figures above are reported by the vendors and community, not our numbers. What we control is the input quality: even light, one background, sharp files.

Next step: decide your tool first, then shoot to its band. Book a slot on the photo studio rental price page and give yourself two hours if you’re building a full set.

What angles and expressions make a strong dataset

Think coverage, not volume. A model learns a face the way you’d learn to recognise a stranger: front-on, in profile, turning, up close, at a distance. Miss an angle and the avatar guesses at it, and its guesses are the part that looks fake.

Here’s the coverage I’d shoot for a personal avatar in a one-hour slot:

• Front, neutral. The anchor shot. Eyes level, relaxed mouth, both ears visible.
• 45 degrees, both sides. Turn the head, not the whole body. This is where most phone sets fail.
• True profile, both sides. Full side-on. Less flattering, very useful to the model.
• Close-up. Head-and-shoulders crop so the face fills the frame at full resolution.
• Wider. Waist-up, so the model reads your build and posture, not just the face.

Then vary the human, not the environment. Neutral expression first, then a light smile, then talking-mid-sentence, then serious. Two or three outfits if your tool allows outfit changes, but keep the styling clean, since busy logos and patterns confuse identity training. Change your face and your clothes. Do not change your light and your wall.

One honest caveat. More expressions help HeyGen-style tools that render new «looks». A tight character LoRA sometimes trains better on a narrower, more consistent set. Match the spread to your tool, and when in doubt, over-shoot angles and under-shoot gimmicks.

Where to go from here: the themed photoshoot sets give you clean backdrops for the outfit variations without dragging your set out of a studio and into a messy location.

Lighting and background: why studio consistency beats a bedroom wall

Consistency is the whole game, and consistency is exactly what a home shoot cannot hold. Move an inch near a window and the light shifts. A lamp throws orange onto one cheek. The wall behind you has a shadow at 3pm that wasn’t there at noon. Every one of those changes is a data point the model has to reconcile, and it reconciles badly.

A studio removes the variables. Soft, even light on the face, held identical across every frame. One clean background, or a cyclorama that reads as flat white or grey with no corner line for the model to latch onto. No hard shadows carving the face into halves the AI then has to average.

That’s why the cyclorama is the quiet workhorse for this. An 8×6-metre curved white wall gives you a background with no edges and no floor seam, so every reference frame shares the same neutral field. The model spends its capacity learning your face instead of learning the geometry of your bedroom.

One boundary worth naming. On our video bookings the two Profoto flashes come free. On a photo-only booking they’re available separately, because a lot of avatar sets get shot on a phone in natural or basic light and that’s genuinely fine for the job. You don’t need a lighting rig to make a clean set. You need even light and a plain wall, and both come with the room.

Before you book: if you want a truly clean, edge-free white or grey field, the cyclorama is the set to book. Message us to check it’s free on your date.

Resolution and technical specs

Straight to the specs, because this is where people quietly sabotage their own set. Shoot sharp, shoot large, and don’t over-process.

• Resolution. Higher is safer. Character LoRAs reportedly want 768px on the short side or better, and 512px is the reported floor. Any modern phone clears this easily. The failure mode is never «too big», it’s motion blur and heavy compression.
• Focus. Every reference frame must be tack-sharp on the eyes. A soft frame teaches soft features.
• No heavy filters. Skip beauty filters, skin smoothing, and aggressive HDR. You’re training the model on a person, and filters teach it a person who doesn’t exist. Retouch later if you must, on the generated output, not the training input.
• Consistent white balance. Lock it. A studio’s fixed light makes this trivial; a home mix of window and lamp makes it a fight.
• File format. Shoot the largest, least-compressed option your device offers. RAW if you can, high-quality JPEG otherwise.

None of this needs a cinema camera. A recent iPhone or Android in a well-lit room clears every bar on this list. The room is doing the heavy lifting.

What to do first: bring the device you actually shoot on, and test one frame in the first five minutes to confirm focus and white balance before you shoot the full set.

Phone vs studio: the hidden cost of bad training data

The reversal most people need to hear: the cheapest set to shoot is the most expensive one to fix. A bad dataset doesn’t announce itself. You upload it, you train, and only when the avatar comes out drifting and plastic do you learn the input was the problem. Now you’re re-shooting, re-training, and burning platform credits on the second attempt.

A muddy training set produces a muddy avatar. Inconsistent light gives you an avatar whose face subtly shifts between clips. Changing backgrounds bleed into the model and show up as artefacts. A soft, filtered set gives you a soft, uncanny face that no amount of prompting fixes, because the fix has to happen at the input.

Here’s the honest maths. An hour of clean studio time is 350 AED. Two hours, enough to shoot a full multi-angle set with outfit changes, is 700 AED. A re-shoot after a failed set costs you that again, plus the training credits you already spent, plus the days you lost. Getting the input right once is the cheapest thing in this whole workflow.

You can absolutely shoot a passable set at home. Plenty of people do. But if the avatar is going on your brand, in your ads, or on a client’s screen, the studio removes the one failure that costs the most to undo.

Next step: if this is a serious avatar for a personal brand or an e-com line, book two hours and shoot it clean once. See the AI photoshoot studio guide for how the clean-input principle carries across AI editing too.

Where SkyLight fits, and where it doesn’t

Plainly: we’re the room, not the software. SkyLight has run self-service studio rental in Dubai Investment Park 2 (SP Warehouses) since 2020, with 7 sets, a cyclorama, and a 4.6-star rating across roughly 290 reviews. We’re open 10:00 to 22:00 every day, minimum booking one hour, prices plus 5% VAT.

What we do: hand you a clean, evenly lit studio and get out of your way while you shoot your own training set. What we don’t do: train your avatar, run your LoRA, generate images, or edit your output. If you want the actual production done for you, that’s a different service and a different company. This page is about renting the space to shoot the input yourself.

Next step: message us on WhatsApp at +971 56 839 9199 with your date and how many hours you need, and we’ll confirm the set.

FAQ

How many photos do I need to train an AI avatar?
Most workflows use 10 to 30 clean, consistent photos of the same person. HeyGen’s Personal Model reportedly recommends 30 or more; character LoRAs reportedly train well on a curated 15-30, where 20 excellent shots beat 200 mediocre ones. Clean beats plentiful.

Can I just use phone selfies for the training set?
You can, and a recent phone in even light is genuinely fine. The problem is consistency: selfies shot at different times, in different light, against different walls, teach the model noise. A studio removes those variables so the model learns your face, not your bedroom.

Do you train the AI avatar for me at SkyLight?
No. We rent you the studio, self-service. You shoot your own reference set and take the files. We don’t train the model, run the LoRA, or generate any images. The room is ours; the dataset and the avatar are yours.

What angles should the training photos cover?
Front neutral, 45 degrees on both sides, true profile on both sides, one close-up, and one wider waist-up frame. Vary your expression and outfit across those angles, but keep the light and background identical the whole time.

What resolution do the photos need to be?
Sharp and large. Character LoRAs reportedly want at least 512px on the short side, with 768px or higher preferred. Any modern phone clears this. The real killers are blur, heavy compression, and beauty filters, not file size.

How long does it take to shoot a full training set?
About one to two hours. One hour covers a solid single-outfit set of angles and expressions. Two hours gives you room for two or three outfit changes and a wider spread, which is why the 700 AED two-hour slot is popular for this.

Why is a cyclorama good for avatar training photos?
An 8×6-metre cyclorama gives a seamless white or grey background with no corner or floor seam. Every frame shares the same clean field, so the model spends its capacity learning your face instead of the geometry of the room behind you.

How much does the studio cost for an avatar shoot?
From 350 AED per hour for photo, or 700 AED for two hours, plus 5% VAT, minimum one hour. Video bookings are 750 AED per hour and include two Profoto flashes; on photo-only bookings the flashes are available separately. Message +971 56 839 9199 to book.