Encinasola

LoRA model trained on images of Encinasola, a village in the province of Huelva, Andalusia, Spain.

Training Data

The model was fine-tuned on a curated set of 24 photographs of Encinasola, including aerial views, landmarks, streets, and landscapes. The training images are not included in this repository. Captions were provided via .txt sidecar files with a 5% caption dropout rate during training.

Training Details

Parameter	Value
Base model	ostris/Flex.1-alpha (Flex.1 architecture)
Method	LoRA (Low-Rank Adaptation)
LoRA rank	linear: 32, conv: 16
LoRA alpha	linear: 32, conv: 16
Optimizer	AdamW 8-bit (adamw8bit, weight decay 1e-4)
Learning rate	1e-4
Training steps	3,000
Batch size	1 (gradient accumulation: 1)
Precision	bf16
Noise scheduler	FlowMatch (sigmoid timesteps)
Loss	MSE
Content/style balance	Balanced
Gradient checkpointing	Enabled
Quantization	qfloat8 (UNet and text encoder)
Resolution	512, 768, 1024 (multi-resolution)
Train UNet	Yes
Train text encoder	No
Save format	Diffusers (bf16), every 250 steps

Samples were generated every 250 steps using FlowMatch sampling (25 steps, guidance scale 4) to monitor training progress.

Usage

Load the safetensors LoRA weights with your preferred Stable Diffusion pipeline and use the trigger word Encinasola in your prompts.

Files

• Encinasola.safetensors — LoRA weights (225 MB)

License

Please respect the rights of the original image owners. This model is shared for personal and research use.