You train a Soul ID, and training requires a paid plan. Open higgsfield.ai, go to Soul, choose Train new character, upload 20 or more photos of the same person (up to 80), wait a few minutes, name the character, then select it from the Character tab every time you generate.
That is the mechanism. The number that decides whether it works is 20, and the second number is how varied those 20 are.
I have watched people feed five front-facing selfies into this kind of tool, get a face that drifts by generation three, and conclude the model is bad. The model is fine. Five photos from one angle is one data point.
If you are going to test this, a referral link carries up to 50% off for three hours: start with the discount applied. We may earn a commission if you subscribe through it, at no extra cost to you.
The 20 photo floor
Soul ID is an identity layer, not a filter. You give it photos of one person, it learns that face, and it applies that face to any generation regardless of preset, lighting, angle or prompt. The docs put the floor at 20 photos and the ceiling at 80, and they are explicit that training quality determines output quality: a smaller set of clean, varied photos beats a large set of inconsistent ones.
The training set is also permanent in a way people do not expect. The character is reusable but not exportable. It lives inside the Higgsfield ecosystem, works across the Soul family, and cannot be downloaded as a file.
Reference image vs trained identity
Both features exist in the same product and they solve different problems.
| Approach | What it does | Where it breaks |
|---|---|---|
| Reference image per generation | Anchors a single generation to a look | Drifts across separate generations, and attaching one disables the prompt field |
| Trained Soul ID | Holds one identity across every generation, style and angle | Not pixel-identical, small drift on extreme styles or unusual angles |
A reference image is a note you leave for one render. Soul ID is training a model on your face once. If you need the same person in twenty images across four outfits and three locations, the reference workflow is the expensive way to get an inconsistent result.
One more limit: a Soul ID holds one person. For a scene with two or more consistent characters, the docs point you to Elements instead.
The training set that produces a usable face
This is the difference between a character that holds and one that melts, straight from Higgsfield's own guidance.
| Helps | Hurts |
|---|---|
| Varied angles and expressions | The same pose repeated across all photos |
| A clear, well-lit face | Heavy shadows or harsh lighting |
| Recent photos, last 4 to 5 months | Outdated photos if appearance has changed |
| At least one full-height shot | Head shots only |
| No obstructions | Sunglasses, hats or scarves over the face |
| One subject per photo | Group shots |
Two entries on the right cause most of the failures. Head shots only means the model never learns your proportions, so a full-body prompt invents a body. Outdated photos means you get a face you had two years ago attached to a style test you are running today.
Higgsfield accepts 20 to 80 photos per Soul ID. Twenty is a floor to clear, not a target. Add angles, not duplicates.
Data table
| Item | Value | Share |
|---|---|---|
| Minimum | 20 | 20.0% |
| Maximum | 80 | 80.0% |
What it costs and where the character can go
Soul ID is part of the Soul family, so the cost question is your plan, not an add-on.
| Plan | Price | Monthly credits | Soul ID training |
|---|---|---|---|
| Free | $0 | 0 | Not available |
| Starter | $19/mo | 270 | Requires a paid plan |
| Plus | $47/mo annual, $59 monthly | 1,200 | Included |
| Ultra | $99/mo annual, $129 monthly | 3,000 | Included |
The docs state that training a character and generating with one both require a paid plan. If your plan lapses, your characters stay saved and become usable again when you resubscribe, which is the friendliest detail in the whole flow.
The generation settings are narrow enough to memorize. Batch size runs 1 to 4 images, quality is 1.5k or 2k, and aspect ratio covers 9:16, 3:4, 2:3, 1:1, 4:3, 16:9 and 3:2. Some plans include a one-time pool of free Soul 2.0 and Soul Cinema generations that does not renew, and those free generations only count on the website, never through MCP or the API.
Where the character can travel matters more than people realize. A trained Soul ID works across Soul, Soul 2.0 and Soul Cinema, and it carries into Seedance video generation as an Element, which appears automatically. That is the path from still images of yourself to video of yourself, and it is the reason to train the identity properly the first time.
The bottom line: twenty clean, varied photos and a paid plan get you a reusable face. Five selfies get you a drift problem you will blame on the model.