Every custom style is built from a combination of AI models that work together to create your desired visual appearance. This guide explains the technical components.

Model Stack Components

A model stack consists of three types of components:

1. Checkpoint (Required) - The base AI model
2. LoRAs (Optional) - Fine-tuning models that add specific features
3. Embeddings (Optional) - Textual inversions that enhance specific concepts

Checkpoints (Base Models)

The checkpoint is the foundation of your style - the primary AI model that generates your images.

Available Architectures

Flux Models - Modern high-quality image generation:

• flux1d - Flux 1 Dev (highest quality, more credits)
• flux1s - Flux 1 Schnell (faster, fewer credits)
• fluxkontextpro - Flux Kontext Pro (balanced quality)
• fluxkontextmax - Flux Kontext Max (enhanced detail)
• fluxkontextdev - Flux Kontext Dev (experimental features)

HiDream Models - Specialized for specific use cases:

• hidreamdev - HiDream Development (optimized styles)
• hidreamfast - HiDream Fast (speed-optimized, lower cost)

Choosing a Checkpoint

For quality-focused projects:

• Use flux1d for maximum quality
• Best for main characters and key scenes

For development and testing:

• Use flux1s or hidreamfast
• Save credits while experimenting

LoRAs (Low-Rank Adaptation)

LoRAs are specialized fine-tuning models that modify your checkpoint's output without replacing it.

What LoRAs Can Do

• Add artistic styles (watercolor, oil painting, sketch effects)
• Introduce specific features (clothing styles, lighting effects)
• Enhance visual elements (detail, atmosphere, color)
• Fine-tune specific aspects of generation

Using LoRAs Effectively

Compatibility:

• Must match your checkpoint's architecture
• System automatically filters compatible options
• Changing checkpoints may require new LoRAs

Stacking:

• Add multiple LoRAs to one style
• Each has a weight/strength setting (0.3-1.0+)
• Lower weights = subtle, higher weights = strong

Best Practices:

• Start with 0-2 LoRAs for clean results
• Test each individually before combining
• Remove if results look poor

Embeddings (Textual Inversions)

Embeddings are learned representations of specific concepts that enhance the AI's understanding.

What Embeddings Can Do

• Define specific characters or archetypes
• Represent complex visual concepts
• Improve consistency for niche subjects
• Strengthen thematic elements

Using Embeddings

• Choose embeddings matching your story's theme
• Multiple embeddings can work together
• Must match checkpoint architecture
• Work alongside your text descriptions

Architecture Compatibility

Key principle: LoRAs and embeddings must match your checkpoint's architecture.

• Flux checkpoint → Flux-compatible LoRAs and embeddings
• HiDream checkpoint → HiDream-compatible LoRAs and embeddings
• Mixing architectures won't work

The system automatically filters and shows only compatible options.

Changing Checkpoints

If you switch to a different architecture:

• Incompatible LoRAs and embeddings are removed
• You'll need to select new compatible models
• Save changes to update the style

Cost Calculation

Credit cost per generation is primarily determined by:

Primary factor:

• Checkpoint model (biggest impact)
• Higher quality = more credits

Secondary factors:

• Number and type of LoRAs (minimal impact)
• Image resolution settings

Not affected by:

• Number of embeddings
• Style name or metadata

Processing Flow

When generating an image:

1. Checkpoint generates the base image
2. LoRAs modify and enhance the output
3. Embeddings refine AI understanding
4. Final image is produced

Technical Limitations

Per-style limits:

• One checkpoint required
• Multiple LoRAs allowed (typically 0-3)
• Multiple embeddings allowed (typically 0-2)

Architecture constraints:

• All components must share architecture
• Cannot mix Flux and HiDream in one style

Next Steps

• Learn how to Create Custom Styles
• Explore Applying Styles to your story
• Review Tips for optimal configurations