Tips and tricks

To train a consistent character in Higgsfield Soul (Soul ID feature for personalization and identity locking), you need to upload 10–50 high-quality photos of the same person/subject from multiple angles, expressions, lighting variations, and poses. The system builds a digital twin/anchor from these, which then carries over to new generations regardless of style, scene, outfit, or prompt.

The goal is dimensionality — give the AI enough data so it understands the full 3D structure of the face, body proportions, hair behavior, posture tendencies, etc. — not just a single flat view.

Recommended Poses / Reference Types for Training Soul ID

Aim for variety in these categories (upload at least 1–3 from each group, ideally 20+ total):

1. Core Face Angles (must-have, 4–8 shots minimum)
   • Front straight-on, neutral expression (eyes open, relaxed mouth)
   • Front straight-on, slight smile / confident expression
   • ¾ left (45° left)
   • ¾ right (45° right)
   • True profile left (90° left)
   • True profile right (90° right)
   • Slight up-angle (looking slightly up)
   • Slight down-angle (looking slightly down)
2. Expression & Emotion Variations (helps with dynamic generations)
   • Neutral / resting face
   • Soft smile / warm expression
   • Confident / slight smirk
   • Serious / focused look
   • Laughing / open-mouth smile (if possible without too much distortion)
3. Body & Pose Variety (critical for full-body consistency)
   • Standing neutral front (arms relaxed)
   • Standing neutral profile (side view, arms relaxed)
   • Standing neutral back view
   • ¾ front standing (weight shift to one hip for natural posture)
   • Seated neutral (torso upright)
   • Arms crossed or hands on hips (shows shoulder/arm shape)
   • Slight action pose (e.g., walking toward camera, reaching up, or squatting lightly — avoids extreme deformation)
4. Lighting & Environment Variations (improves robustness)
   • Natural daylight (window light)
   • Indoor soft overhead light
   • Slightly directional side light (to show face structure)
   • Avoid heavy shadows, extreme low-key, or colored gels
5. Bonus / Advanced (if you have 30–50+ photos)
   • Different hairstyles or slight hair movement (e.g., ponytail vs loose)
   • Different outfits (same person, different clothes — helps with clothing adaptability)
   • Close-up face only (for fine facial detail lock)
   • Full-body distant shot (for overall proportions)

Quick Checklist for Best Results

• Quality first: Well-lit, sharp, high-resolution (at least 1024px on the short side), no motion blur, no heavy filters
• No occlusions: No sunglasses, masks, hands over face, extreme angles that hide features
• Recent photos: Use images from similar time period (same age/weight/hair) for best match
• Diversity over quantity: 15–25 varied good shots > 50 identical front-facing selfies
• Upload order: Doesn’t matter much, but start with strongest front + profiles

Once trained, you can reference the Soul ID in future prompts to generate the same character in any pose, outfit, or scene with high fidelity.

If you upload your training set screenshots or describe what poses you already have, I can suggest which ones to add next for maximum improvement.