The Number One Problem With AI Characters
You generate a perfect image of your AI character. Gorgeous. Exactly the look you want. Then you generate the next image and she looks like a completely different person.
This is the consistency problem, and it kills more AI influencer projects than anything else.
What Soul ID Does
Higgsfield Soul ID locks in a character identity using reference images. Think of it as face recognition in reverse — instead of identifying someone, it generates someone with consistent features.
Consistency Rate
- Without Soul ID: 40-60% consistency
- With Soul ID: 80-90% consistency
- With Soul ID + LoRA: 95%+ consistency
Setup: Reference Image Selection
- Front-facing neutral: Clean, well-lit, no extreme expression
- Three-quarter angle: Shows face structure from the side
- Smiling: Captures how the face changes with expression
- Different lighting: Proves features are consistent across conditions
- Close-up detail: Eyes, nose, mouth at high resolution
Workflow: T2I to I2I Pipeline
Text-to-Image (T2I)
- Write your prompt describing scene, outfit, pose
- Include character description keywords matching references
- Generate with Soul ID active
- Select the best result
Image-to-Image (I2I) Refinement
- Take the T2I output
- Run through I2I with same Soul ID references
- Increase denoising to 0.3-0.5 for subtle corrections
Alternatives to Soul ID
LoRA Training
Train a LoRA on 50+ images. Takes 20 minutes on M4 Ultra. Highest consistency.
IP-Adapter
ComfyUI node using single reference image. Lower consistency but works with any SD model.
InstantID
Zero-shot face transfer. Good for quick results but less controllable.
FAQ
What is Soul ID?
Soul ID is a Higgsfield feature maintaining AI character facial consistency using reference images to constrain generation.
Do I need a GPU for Soul ID?
Soul ID runs on Higgsfield servers. For LoRA training, a Mac Studio or GPU with 16GB+ VRAM is recommended.
How many reference images do I need?
Minimum 3, recommended 5. Diminishing returns past 7-8 images.