| Property | Value |
|---|---|
| License | FAIR AI Public License 1.0 |
| Base Model | Illustrious-xl-early-release-v0 |
| Training Duration | 50.5 hours |
| Hardware Used | 4x RTX3060 |
What is arcain?
Arcain is a specialized LORA (Low-Rank Adaptation) model trained on the Illustrious-xl v0.1 base model. It represents a significant advancement in fine-tuning capabilities, specifically designed to be compatible with various ILXL-based models including NoobAI-XL.
Implementation Details
The model was trained using the kohya-ss/sd-scripts framework with sophisticated parameters including a network dimension of 128 and matching alpha. The training process utilized mixed precision (bf16) and implemented gradient checkpointing for optimal performance.
- Utilized Lion8bit optimizer with a learning rate of 1e-5
- Implemented constant warmup scheduler with 100 warmup steps
- Employed cache latents with disk storage for efficiency
- Training conducted over 8 epochs with checkpoints saved every epoch
Core Capabilities
- Compatible with multiple ILXL-based models
- Optimized for 1024 resolution with bucket range of 512-4096
- Implements caption tag dropout for improved generalization
- Supports efficient batch processing with gradient accumulation
Frequently Asked Questions
Q: What makes this model unique?
The model's unique features include its training configuration using advanced techniques like min_snr_gamma and full_bf16 precision, along with its compatibility across multiple ILXL-based models.
Q: What are the recommended use cases?
This model is particularly suited for applications requiring high-resolution image processing, with optimal performance at 1024 resolution and support for various bucket sizes.