Higgsfield Soul ID: how to keep the same face across images, and the consent rule that comes first

Soul ID is a character feature inside Higgsfield, an AI platform that makes images and video and hosts its own image models called Soul. It lets you save one face as a trained character and call it up for future images. It is useful for a founder who wants consistent branded pictures, or a brand mascot. It is also a tool that makes realistic pictures of a real person, which is why this page puts consent first, before the steps.

This page may contain affiliate links from which Blur Studio earns commissions.

This guide uses Higgsfield's help center (published August 2026, updated September), its Terms of Use (last updated 23 October 2026), and a Higgsfield-sponsored test, which I label. It is information, not legal advice. Check local law and the platform's current rules.

The rule that comes first: consent

Only train a Soul ID on a person who agreed. That means yourself, or someone who has clearly said yes to having their face used this way. Higgsfield's help page says to upload only photos of yourself or of someone who has given you permission to use their likeness. Never train one on a customer, a competitor, an ex-employee, a stranger from social media, or a celebrity.

The Terms back this up. They say that if your content contains the name, face or voice of any person, including you, you must have obtained all the rights, releases and consents needed (section 4.2). For face or voice inputs you must hold those consents and give any disclosures the law requires, and you may not use the service to identify or verify people (section 5.3). Higgsfield does not check this for you. It is your responsibility.

Here is a consent routine I would use. It is a starting point, not a legal form, so ask a lawyer to review it for your country.

1. Get a written release. Name the tool type (AI image and video tools), what the pictures are for (for example, campaign X), how long the permission lasts, and whether the trained character may be reused for later work.
2. Tell the person about training. Higgsfield's Terms let it use inputs and outputs to train its models, except under an Enterprise Agreement. A face photo is an input. The person should know that before agreeing.
3. Record where the photos came from and who took them. You also need the photographer's permission to use the photos.
4. Treat the voice as a second consent. A face does not carry a voice. If you plan to add a voice, that needs its own agreement.
5. Write down how to undo it. If the person leaves or changes their mind, you should know how to delete the character and stop using the pictures.
6. Add a disclosure. Where the law or platform rules require it, say the image is AI-made. For an AI version of a real executive, I would add a caption as a house rule even where it is not required. Higgsfield's Terms also say not to remove the provenance marks it may embed.

Also think about who owns the account. Higgsfield says credits, subscriptions and generations cannot be moved between accounts, and a Soul ID cannot be exported. So train it in the business account, not in a freelancer's personal one. Setup tips are in Higgsfield login and account setup.

What Soul ID is, and what it is not

You upload photos of one person, Higgsfield trains a character, and you pick that character in the Character tab when you generate. Per the help center, a trained Soul ID works across the Soul family of image models (Soul, Soul 2.0 and Soul Cinema). In Soul Cinema, open the Character panel and run the update if it is offered.

What it is not:

• Not a pixel-exact copy. Higgsfield says to expect "clearly the same person".
• Not for two people. One Soul ID holds one person. For two or more consistent characters in one scene, Higgsfield points to Elements (reusable references you can call up in a prompt).
• Not a file you own. It cannot be exported.
• Not a voice. A Higgsfield-sponsored test found that the voice did not carry with the face across tools.

Train a Soul ID in five steps

1. Confirm consent and write down where the photos came from.
2. Open the Soul tool on Higgsfield's website and choose to train a new character. Names of buttons may change.
3. Upload 20 or more photos of the one person. Up to 80 is allowed.
4. Wait a few minutes. Higgsfield's own blog says 3 to 5 minutes while another source says about ten, so go with the help center wording.
5. Name it and select it in the Character tab. Test it before you use it on real work.

Higgsfield's help center and pricing table print no training cost, and some Higgsfield blog posts quote one. I am not repeating a blog figure as a price, so read the cost on the button before you confirm. Check the pricing guide for what plans cost now.

The photo set that works, with a pass or fail checklist

Higgsfield's guidance: varied angles and expressions, clear well-lit faces, recent photos (within the last 4 to 5 months, if looks have changed), at least one full-height photo, no sunglasses, hats or scarves, and one person per photo. It says a smaller set of clean varied photos beats a large messy one. Turn that into a gate. The counts below are my own thresholds, not Higgsfield's:

• At least 20 photos, no more than 80.
• At least 8 different angles or expressions.
• At least 3 lighting conditions.
• At least 1 full-height photo.
• No group shots, no filters, no beauty-retouch apps.
• No photo older than your current look.

Pages you may see online disagree on the minimum: one Higgsfield blog (October 2026) says it accepts as few as 5 photos while recommending 20 to 80. The help center says 20 or more, so I would follow it. That blog also recommends 960 pixels or more per image, so skip tiny screenshots. If you cannot pass every line, do not train yet. Take better photos first. A weak set gives a weak identity, and you pay credits to find that out.

Use it in images, then in video

In stills, pick the character and write a normal prompt. In video, the help center documents one route: the character carries into Seedance (an AI video model) as an Element, which you add by typing @ in the prompt and choosing it from My Elements. A Higgsfield marketing blog dated 27 July 2026 says the same identity carries across Kling 3.0, Veo 3.1, Seedance 2.0 and Wan 2.6 (AI video models from Kuaishou, Google, ByteDance and Alibaba). The help center documents only the Seedance route, so treat it as marketing until you test the exact model you plan to use. Video costs more credits than images, so test on an image first.

One risk with real faces in video. In Cinema Studio (Higgsfield's workspace for film-style shots), Higgsfield's content protection can block a clip with the message "May contain protected content" when it detects a real face or protected material. The help page tells users to remove real faces or recognisable protected material and retry. The help page does not say whether a Soul ID of a real person triggers it, so run a cheap test and keep a fallback, such as an invented character. See the Cinema Studio guide.

Keeping the face steady from clip to clip

A Soul ID covers stills. For a run of video clips, Higgsfield's blog (29 June 2026) suggests stacking a few cheap habits on top. These are vendor tips, not guarantees, so test each one. The advice to keep chains short is my own rule of thumb:

1. Start clips from a picture, not from text. Make the portrait with your Soul ID, upload it as the first frame, and let the prompt describe only the action and the camera move.
2. Paste one fixed character description (hair, clothing, age range, marks) at the start of every prompt.
3. Chain short. Use the last frame of one clip as the first frame of the next. I would go only two or three clips deep, because small errors can pile up, then start a new chain from a fresh Soul ID portrait.
4. Repair, do not redo. The same blog says Higgsfield's Genjutsu (its edit-an-existing-clip tool) can fix a drifted face in a finished clip from reference photos. Test it on one clip before you rely on it.

If the character does not exist as a real person, Higgsfield's route is to make 20 or more portraits of an invented character in its AI Influencer Studio first, then train a Soul ID on them. Check every portrait for matching features before training.

A 30-minute test plan with pass or fail

1. After training, make four images in the cheapest settings: a neutral front portrait, a three-quarter view, a profile, and a low-light shot.
2. Ask two people who know the subject: is it the same person in all four?
3. Pass: both say yes for at least three of the four, and the subject says they are comfortable with the results.
4. Fail: a clear drift in any shot, or any hint of discomfort from the subject. Stop, retrain with a better set, or drop the plan.
5. Only then make one test clip, and run the same two-person check.

Those thresholds are mine, not Higgsfield's.

Why it drifts and how to fix it

Symptom	Likely cause	Try
Face looks off in hard light or with the head turned down	Training set lacks those angles and lighting	A Higgsfield-sponsored test saw a face looking down under harsh light read noticeably older, and left such images out of its training set. Add varied, well-lit photos and retrain.
Looks older or younger than the person	Old or mixed-age photos	Use recent photos only.
A second person's features appear	A group shot or reflection in the set	Remove photos with other faces.
Small details change (freckles, hair tie, shirt)	Normal model drift	Describe key details in the prompt and re-run.
Stylised or cartoon outputs lose the face	Extreme styles	Move toward a photo-real look or accept weaker likeness.

If you train an invented character from generated portraits, check every training picture before you use it. Only the first row above draws on a published test. The other causes are my judgment.

Limits to tell clients

• Close likeness, not identical.
• It lives inside your Higgsfield account.
• It is deleted with the account. Higgsfield says deleting an account removes trained Soul characters after a 30-day window. Content already used for training cannot be pulled back.
• You need a retraining plan if the person's look changes.

Keep a record per character: client and person, date, version (for example, acme-founder-v1), the photo set, and a contact sheet of approved results. Because the trained identity cannot be exported, your photo set and approved outputs are your backup.

Privacy, training and deletion

Higgsfield's Terms say it may use your content, including inputs and outputs, to train and improve its models, except under an Enterprise Agreement. Deleting content or the account stops future use, but content already trained on cannot be removed. Its privacy policy says it does not store biometric information from face or voice features, which it uses only transiently to provide the feature; read the policy yourself before uploading a client's staff. Because the photos are an input, tell the subject before they agree. If that is a deal breaker, ask Higgsfield about its Enterprise route. More on the company is in is Higgsfield safe and legit.

Soul ID vs Elements vs LoRA: pick the right one

Option	What it is	Best for
Soul ID	A trained face inside Higgsfield	Many still images of one real or invented person
Elements	Reusable references you call up in a prompt, created from images	Quick starts, fictional characters, two or more characters in a scene, video
A single reference image	One picture used per generation, as many image tools allow	A one-off shot
A LoRA	A small add-on you train and run on your own machine, for example in ComfyUI (a free image workflow tool)	An identity file you control, offline and portable

The author of the Higgsfield-sponsored test concluded that an Element is the better place to start, and that a Soul ID suits still-heavy work. I agree with starting cheap.

Who should skip Soul ID

• You need one throwaway image. Re-prompting may be quicker.
• You cannot get the person's agreement in writing.
• You want a portable file or full offline control.
• The person is not you or a consenting spokesperson. A customer, a rival and a celebrity are never fair game.

My ranking of sound uses: your own face for personal branding, then a consenting founder or spokesperson for a defined campaign, then an invented brand mascot trained on generated portraits.

How I'd approach this

Get the consent first, in writing. Train in the business account, never a personal one. Pass the photo checklist, run the 30-minute test, and spend video credits only after the stills pass. You can start from the Higgsfield site if you decide it fits. If you need a record of consents, character versions and approvals kept in one place, Blur Studio builds approval portals and shared asset libraries around tools like this: see how that works.