What is Luminia-13B-v3?
Nekochu/Luminia-13B-v3 is a 13 billion parameter language model, fine-tuned by Nekochu from the Llama-2-13b-chat base. This iteration, continuing from Luminia-13B-v2-QLora, is specifically engineered to generate detailed Stable Diffusion prompt metadata from concise English descriptions. It leverages its reasoning capabilities to expand short summaries into comprehensive prompts suitable for image generation.
Key Capabilities
- Stable Diffusion Prompt Generation: Transforms brief textual inputs into rich, descriptive prompts for Stable Diffusion models.
- Reasoning for Prompt Enhancement: Utilizes advanced reasoning to infer and expand upon user descriptions, creating more effective image generation prompts.
- Flexible Output: Capable of generating a wide range of prompts, including those for NSFW content, based on the input.
- Quantization Support: Available with various quantization options including exllamav2 (2.4bpw-h6, 4.25bpw-h6, 8.0bpw-h8) and GGUF (Q4_K_M, IQ4_NL) for optimized deployment.
Training Details
Luminia-13B-v3 was trained using QLoRA on a single GPU with 24GB VRAM, utilizing the hiyouga/LLaMA-Efficient-Tuning framework. Key hyperparameters included a learning rate of 5e-05, a cutoff length of 4096, and a LoRA rank of 32, targeting all layers. The model uses the Alpaca prompt template.
Should I use this for my use case?
This model is ideal for developers and creators who need to automate or enhance the creation of Stable Diffusion prompts. If your application involves generating diverse and detailed image prompts from minimal input, especially for creative or artistic endeavors, Luminia-13B-v3 is a strong candidate. Its ability to handle both SFW and NSFW content provides versatility for various creative projects.