Why Character Consistency Is a Production Problem First
The reason a recurring character is hard in AI video is not that the models cannot generate a convincing face - they can. The problem is that 'convincing' and 'the same person' are two different outputs. Without a persistent identity anchor, a model that generates a great face in clip one will generate a similar but subtly different face in clip two, because every generation is a fresh interpretation of whatever description or reference it was given.
This drift becomes visible at scale. A single clip looks fine. Ten clips in a series where the character ages slightly between episodes, changes bone structure across a lighting shift, or looks like a sibling instead of the same person - that is a consistency failure the audience notices even if they cannot name it. Soul ID solves the front half of this problem: the identity holds because it was trained and stored, not re-described each time.
Building the Training Set That Actually Holds
The training set is the most important decision in a Soul ID workflow. A weak photo set produces a weak identity, and a weak identity drifts under style pressure. Before uploading anything, audit the photos against this checklist.
- 20 photos minimum, up to 80. More is not automatically better - 30 clean and varied photos outperform 80 repetitive ones.
- Include at least one full-height photo, not headshots only. The model needs body language and proportion data, not just the face.
- Cover multiple angles: front, three-quarter, slight profile. A training set that is all straight-on shots produces an identity that softens at angles.
- Vary lighting conditions. Indoor, outdoor, different times of day. The identity should hold under a cinematic key light and under a soft overcast sky.
- No sunglasses, hair covering the face, heavy shadow across the eyes, or group shots. One clean subject per photo, full stop.
- Use recent photos - within four to five months - so the trained identity matches how the person looks now, not three years ago.
The Prompt Structure That Keeps the Identity Locked
Training the Soul ID is only half the workflow. How you prompt with it determines how much of the trained identity survives style pressure. A prompt that describes the character from scratch in addition to calling the trained ID creates a conflict: the model tries to satisfy both the trained identity and the new description, and they pull in different directions.
The right prompt structure calls the Soul ID through Elements using @ and then describes the scene, lighting, wardrobe, and camera direction - not the person. Let the trained identity handle who the person is. Your prompt handles everything else.
Prompt structure: what to specify and what to leave to the trained ID
| What the prompt should specify | What to leave to the trained Soul ID |
|---|---|
| Scene location and environment | Face structure and identity |
| Wardrobe and styling | Skin tone and texture |
| Camera motion and framing | Facial features and expressions |
| Lighting setup (key light, backlight, ambient) | Age and proportion |
| Mood and color grading direction | Eye color and bone structure |
Running a Content Series: The Production Workflow
A content series with a recurring character looks like this in practice. Train the identity once from a strong photo set. Name it something descriptive - the character's name and role, not just a number. Then run a test batch of 5 to 10 clips across different scenes, outfits, and lighting setups before committing to the full series. The test batch is where you find whether the training held.
- Train the Soul ID from a strong, varied photo set. Run the training, name the character.
- Generate a 10-clip test batch: vary scene, wardrobe, lighting, and camera framing. All scenes, no same setting twice.
- Review the test batch as a group, not individually. Look for drift in bone structure, skin tone, and proportion across the range.
- If drift is within acceptable range, lock the prompt template and move to production. If not, audit the photo set first, then retrain.
- For the production run, batch-generate all clips in a series before editing any of them. Reviewing all clips together catches consistency failures that are invisible when you review one at a time.
- In final editing, treat minor inter-clip variation as normal. Color grade the series uniformly so the character's palette stays consistent even if small generation differences exist.
Multi-Character Scenes: Stacking Elements
One Soul ID holds one person. For scenes with two or more consistent characters, train each identity separately and then stack them in a single prompt using Elements. Each character gets a separate @ call in the prompt, and you describe the spatial relationship between them - who is in the foreground, who is in the background, how they are positioned - rather than describing either character's appearance.
Multi-character scene generation is harder to hold consistent than single-character. Identity drift compounds across two trained IDs in the same frame. The practical limit for reliable results is two characters per scene. Three or more starts requiring significantly more prompt engineering to hold stable.
Client Work: Using Soul ID on the Client's Own Likeness
Client work is one of the highest-value applications of Soul ID, because it solves a real problem that clients pay to solve: a consistent face across a campaign without booking a model for 12 separate shooting days. A fashion brand that wants the same person across a spring campaign can get that from a trained Soul ID on a model they already have approved photos of.
The mechanics are the same as a personal content series - strong training set, scene-only prompting, batch review. The additional considerations for client work are consent and legal clarity. The Soul ID is being trained on a real person's identity. The client needs to supply photos of someone whose likeness they have explicit rights to use, and that agreement should be documented in your client contract before training anything.
The pricing angle: charging per deliverable without disclosing that you are using AI is one model, and disclosure requirements vary by client and contract. Charging as a retainer for AI video production across a campaign is another. The retainer model scales better, because a trained Soul ID means the marginal cost of additional clips drops significantly after the training is done. See the guide on [how to sell AI video as a monthly retainer](/blog/how-to-sell-ai-video-as-a-monthly-retainer) for how to structure that conversation.
The Higgsfield Income Club Workflow Template
The full Identity Anchor Kit - which covers how the manual method and Soul ID work together for a complete consistency workflow - is inside the Higgsfield Income Club at higgsfieldincomeclub.com for $9 a month. The kit includes the photo-set audit template, the locked prompt structure, and the batch review checklist that members use to run client campaigns and personal content series at volume.
Soul ID is the right tool once you know you are building a real recurring character - a series that will run for months, a client campaign with a defined look, or an AI influencer account where consistency is the product. The [Soul ID explainer](/blog/higgsfield-soul-id-explained) covers the feature itself. This guide covers using it as a production system.
Short, practical drops on income paths, prompts that work, packaging for views, and getting paid. No spam, unsubscribe anytime.
Frequently asked questions
How is Soul ID different from just using a reference image?
A reference image re-anchors a single generation and tends to drift across a batch or a series, because you re-supply it from scratch each time and the model reinterprets it slightly differently. Soul ID trains a reusable identity that the model holds across every generation in every session without re-uploading anything. The consistency is stored, not manually applied.
How many photos do I need for a good training set?
20 to 30 clean, varied photos - different angles, lighting conditions, and at least one full-height shot - produce better results than a larger set of inconsistent or repetitive ones. 80 is the platform maximum, but quantity does not substitute for variety and quality.
Can I use Soul ID for an invented character, not a real person?
Soul ID is built to train on real photos of a real person. For an invented character, design the look first using the manual Identity Anchor Kit method - a locked reference image set plus a written character description block - and then train a Soul ID on the generated portraits of that character if you want to lock it as a platform-wide asset.
Does Soul ID work in Kling, Seedance, and other models inside Higgsfield?
Yes, through Elements. A trained Soul character appears in Elements automatically once training is complete, and you can pull it into any model that supports Elements by using @ in the prompt field.
How do I run a multi-character scene with Soul ID?
Train each character's identity separately as its own Soul ID. In the prompt, call both characters using @ Elements and describe the spatial relationship between them. Do not describe either character's appearance in the prompt - let the trained identities handle that. Batch generate multi-character scenes separately from single-character scenes and review them as a group.
What do I do if my Soul ID generations still drift after training?
Audit the training photo set first. Drift after training is almost always a photo-set problem: too few photos, insufficient angle variety, inconsistent lighting, or including photos where the face is partially obscured. Retrain from a stronger set before changing your prompt structure.
Can I train a Soul ID on a client's likeness for commercial work?
Technically yes, but only with clear documented consent from the person whose likeness is being trained. Include explicit likeness usage rights in your client contract before training anything. This locks a real identity into a generation tool, and that requires a clear paper trail.
Last reviewed by David on September 12, 2026
