Project Idea: To create a visually cohesive herbarium of non-existent plants in the style of classic archival botanical illustration. The LoRA is trained on a proprietary image collection and is then used to generate new species not present in the dataset.
Final series—six botanical illustrations of the fictional plant NEOFLORA, generated using a trained SDXL LoRA after training on a dataset of 30 flower photographs. The project establishes a cohesive visual concept of an «unknown flowering species» rendered in the style of archival botanical illustration. The source photos were captioned using BLIP with the prefix «botanical illustration of NEOFLORA.» The model’s objective was to internalize the generalized morphology, the scientific precision of the drawing, and the soft lighting. The series tests the concept’s consistency across various environments:
- Giant nocturnal flower in a dark forest
- Aquatic plant on a lake
- Desert variant with seed pods
- Climbing plant on an abandoned building
- Microscopic scientific specimen
- Sample in a futuristic greenhouse
The model maintains the recognizable complex morphology, the academic style of the illustration, and the feel of a museum specimen. While the scale, environment, and growth habit change, the «family» resemblance remains consistent.
The latest version of ChatGPT was utilized in the project to resolve library dependency conflicts within the project.