Make photorealistic AI photos that actually look like you
Most AI photo tools give you a good-looking stranger. You upload one selfie, and what comes back has your haircut and somebody else’s jaw. Higgsfield’s fix is to stop treating your face as a reference image and train it as a character you can cast in any shot — a Soul ID. You do the upload once, it trains in a few minutes, and after that every image you generate is unmistakably you.
Step by step
Understand what you’re actually making
Start at higgsfield.ai/soul-intro to see what the model does. Soul 2.0 is Higgsfield’s photo model — it’s tuned for editorial, shot-on-a-camera realism rather than glossy AI gloss, and it ships with 20+ built-in style presets so you don’t have to write paragraphs of camera jargon.
Soul ID is the part that matters here. It trains a small model on photos of one person, so your likeness is baked in rather than guessed at from a single reference. Once trained, the same face holds across every pose, outfit, lighting setup and preset you throw at it. Training takes roughly 3–5 minutes and costs around 25 credits, and Soul ID needs a paid Higgsfield plan.
Do the character first. Everything else in this guide depends on it existing.
Gather 20+ photos that describe your whole face
This step decides your result. Go to higgsfield.ai/character and choose to train a new character — you’ll need at least 20 photos, and it accepts up to 80.
What you’re assembling is a description of your face from every side, not a highlight reel. Include straight-on shots, three-quarter angles and profiles. Vary your expression a little — neutral, smiling, mouth closed, mouth open. Keep the lighting clean and fairly consistent across the set (soft daylight is ideal), aim for at least 960px, and include at least one full-length shot so it learns your build and proportions, not just your head.
Use recent photos — the last four or five months. If your hair, weight or facial hair has changed, older photos will drag the likeness back toward a version of you that no longer exists.
Now the ruthless part. Cut anything with sunglasses, a hat shadowing your eyes, a scarf across your face, or heavy shadow. Cut group shots and anything where you’re not clearly the subject. Cut cropped faces. Cut anything with a beauty filter or skin smoothing — filtered photos teach it a fake version of your face, and you’ll see that plastic quality in every image afterwards. And cut near-duplicates: ten photos from the same afternoon in the same top count as roughly one photo.
A tight set of 25 varied, honest photos beats 70 inconsistent ones every time.
Train it and give it a name
Upload the set, give the character a clear name you’ll recognise in a dropdown later (your own name is fine), and start training. It takes a few minutes.
You only do this once. The character is saved to your account and reusable forever — you never re-upload those photos again.
Open Soul 2.0 and cast your character
Go to higgsfield.ai/ai/image?model=soul-v2. Two things to check before you type anything:
One — the model is Soul 2.0. Higgsfield hosts several image models and the picker doesn’t always land where you expect. If you generate on the wrong model, your character either won’t be available or won’t be applied.
Two — your character is selected. Open the Character tab and pick the one you just trained. This is the step people skip; without it you get a beautiful photo of somebody who isn’t you.
Then set your frame: Soul supports 1:1, 4:5, 9:16, 16:9, 3:4 and 4:3, and generates up to 4K natively — so pick the shape for where it’s going (9:16 for stories and reels, 4:5 for an Instagram post, 16:9 for a website header) rather than cropping later.
Write the shot, not the face
The instinct is to describe yourself. Don’t — the character already handles who is in the frame, and piling on facial description just fights it. Your prompt’s job is everything around the face: wardrobe, location, light, camera, mood.
Pick a preset first and let it do the heavy lifting — Warm Ambient, Subtle Flash, Editorial Street Style, Flash Editorial and the rest are pre-tuned looks, and choosing one is faster and more consistent than describing lighting from scratch. Then add your specifics in the prompt.
Realism lives in imperfection. Ask for real fabric texture, visible skin texture and pores, natural window light, a little grain — and explicitly rule out the tells that make an image read as AI: airbrushed skin, plastic sheen, over-symmetrical faces, glowing rim light, that waxy studio look.
Generate in batches, then judge them cold
AI image generation is a slot machine — the same prompt gives a different result every time. Generate several, walk away for a minute, and come back to pick. You’ll judge your own face more honestly when you’re not watching them render.
Zoom in on the usual failure points: hands, teeth, ears, jewellery, glasses frames, text on clothing, and the transition between hair and background. Then check the thing only you can check — does it look like you, or like your slightly-symmetrical cousin? If several images drift the same direction, that’s your training set talking, not the prompt.
When one preset is landing, stay with it for a whole set. Switching presets every generation is how you end up with twelve good photos that don’t look like they belong to the same person.
Reuse the character everywhere else
Your character isn’t locked to one page. Generate an image you like, open it in Assets, and choose Create Element from its menu. That turns your likeness into an Element you can call with @ in any prompt field — including Higgsfield’s video models.
That’s the real payoff: one trained face, reused across headshots, thumbnails, campaign images and video, all recognisably the same person.
Retrain when you change
A Soul ID is a snapshot of how you look right now. New haircut, new glasses, a year of ageing — the likeness quietly drifts out of date. When that happens, train a fresh character on new photos rather than fighting it in the prompt.
Keep the old one. Having “me, 2026” alongside “me, long hair” is genuinely useful, and training is cheap compared to the time you’ll waste re-rolling a character that no longer matches your face.