Slay The Spire 2 - Flux.2 Klein 9b style LORAs
Reverse-engineered 655 game images to create a dark fantasy style model for AI art generation.
Independent AI art creator Dever has released a new style LORA (Low-Rank Adaptation) for the Flux.2 Klein 9b distilled image model, trained directly on the extracted artwork from the highly anticipated game Slay the Spire 2. Using GDRE Tools to reverse-engineer the game's files, Dever compiled a dataset of approximately 55 event illustrations and 600 card images, totaling 655 pieces of original game art. He trained two variants of the model: one on the event art only and another on the full combined dataset. The resulting LORA allows users to generate new images or edit existing ones in the exact dark fantasy illustration style of the game by using the trigger word `sts2_style`.
The model is hosted on Hugging Face and is designed for use with the Flux.2 Klein 9b base model for text-to-image generation. Dever notes the training included some of the game's unfinished work and sketches, providing a unique glimpse into the artistic process. He explicitly states the model is for non-commercial use only due to being trained on copyrighted material. In a community-focused move, Dever has also offered to potentially train models for others who have consistent style datasets but lack the GPU resources, inviting them to send samples via direct message.
- Trained on 655 images (55 events, 600 cards) extracted from Slay the Spire 2 using GDRE Tools.
- Two model variants released: one trained on event art only, another on the full combined dataset.
- Uses trigger word `sts2_style` and is for non-commercial use with the Flux.2 Klein 9b distilled model.
Why It Matters
Demonstrates how AI tools can replicate specific, high-value art styles, raising new questions about copyright and creative access in game development.