Lugubriate (Scribble Art) Style LoRA for Qwen 2512

Independent AI artist and developer ThePoetPyronius has launched a new style-specific LoRA (Low-Rank Adaptation) for the Qwen 2512 image generation model, dubbed 'Lugubriate.' This model fine-tunes Qwen 2512 to produce images in a distinct monochrome, black-and-white 'scribble art' style, inspired by the emergent art technique where coherent forms are drawn from random scribbles. The creator describes it as having a 'creepypasta' and 'dank' vibe, aiming to capture beauty born from visual chaos.

The technical training was conducted using the Ostris AI Toolkit. The LoRA was trained over 25 epochs and 6,000 steps on a dataset of 80 curated images. Key parameters include a rank of 16, BF16 precision, and the use of an 8-bit transformer and text encoder. The creator notes the LoRA works best at a strength of 1.0 but can be dialed down to 0.8 for 'softer edges' and to encourage interesting 'style bleed' and subtle coloring effects, offering users creative control over the final aesthetic output.

This release highlights the continued trend of niche, community-driven model fine-tuning. By sharing the complete training specs—including the AdamW8Bit optimizer, a learning rate of 0.0003, and the use of Differential Guidance—ThePoetPyronius provides a valuable case study for other hobbyists looking to create their own highly specific artistic styles, democratizing advanced model customization beyond major labs.