It all started when I was struggling to pick between different climbers for my balcony: I grew frustrated with the approximations of most mainstream models when it came to specific plants, so I decided to make my own. One thing led to another and here I was, building ComfyUI workflows so I could get usable and clean 3D assets from these plants.

The dataset is built from CC0 images (iNaturalist, Wikimedia Commons), checked one by one. I wrote a few captions by hand to fix a controlled vocabulary (species, season, coverage, wall, framing), then had an agent caption the rest from those examples, and spot-checked them. The same words later work as the interface for generating assets.

I started with a single SDXL LoRA for all species, then tried DoRA. Both mixed up wisteria and clematis, which shared too much vocabulary. One model per species solved it. Moving to FLUX.1 [dev] gave much more accurate leaves and flowers, and LoKr learned the shape of individual flowers better than LoRA, at the cost of larger files and slower training. Each comparison changes one thing only, with fixed prompts and seeds.

Prompts can’t keep a building identical from one season to the next, so in ComfyUI a precise inpainting mask is necessary. The facade stays identical to the pixel, and the denoise strength can work as a season slider.

A small MCP server wraps the pipeline as tools an AI assistant can call: generation, inpainting etc. It holds the right model, the validated settings and the training vocabulary, so the user only describes what they want. The server was built with Claude Code: I designed the tools and the pipeline, and have been adding species myself once their models are trained and uploaded to Hugging Face.

FLUX.1 [dev] is under a non-commercial licence: the Flux models and images from this project are for demonstration.

Models and model cards: huggingface.co/Potiron
MCP server: github.com/MajesticLog/botanical-flux-mcp