clay-teapot-flux-lora

Clay Teapot FLUX LoRA

This repository contains a LoRA fine-tuned from the gated FLUX.1-dev base model to generate product-style Yixing zisha clay teapots.

Files

The checkpoint is 171,969,416 bytes and is the final artifact of the dated 700-step training run.

Usage

You must separately obtain access to and accept the license for black-forest-labs/FLUX.1-dev; this repository does not include base-model weights. The reproducible AI Toolkit inference command and training setup are documented in the project repository.

The manually validated generation settings are FLUX FlowMatch at 768×768, 20 inference steps, guidance scale 3.5, BF16, and the low-VRAM quantized path.

Training data and limitations

Training data is available from the public lyuxinghe/clay-teapot dataset. One manually reviewed inference from this checkpoint produced a sharp, recognizable teapot. No automated image-quality benchmark or comprehensive held-out prompt-fidelity evaluation is provided; inspect outputs before using them in downstream work.