The model was adapted using the Smithsonian Butterflies dataset to generate butterfly-style images using a Denoising Diffusion Probabilistic Model (DDPM).

Training Details

Base Model:
- google/ddpm-celebahq-256
Dataset:
- huggan/smithsonian_butterflies_subset
Resolution:
- 256x256
Training Method:
- Full U-Net fine tuning
Epochs:
- 15
Batch Size:
- 4
Learning Rate:
- 1e-05
Optimizer:
- AdamW
Mixed Precision:
- FP16

Features

✅ Fine tuned diffusion model
✅ Supports Hugging Face Diffusers pipeline
✅ Faster sampling using DDIM scheduler
✅ Checkpoint based training
✅ Gradient checkpointing enabled

Usage

python

from diffusers import DDPMPipeline


pipe = DDPMPipeline.from_pretrained(
    "tenperformer/ddpm-butterfly-finetuned"
)

image = pipe(
    num_inference_steps=40
).images[0]

image.show()
Training Notes

The model was trained with:

Gradient accumulation
Mixed precision training
Periodic checkpoints
Hugging Face Hub backup
Limitations

This is an experimental fine-tuned DDPM model.

Outputs may contain artifacts due to:

Small dataset size
Limited training duration
DDPM sampling randomness
License

MIT


![image](https://cdn-uploads.huggingface.co/production/uploads/6a0cd36102135372849ae9d0/ljnHPPqICVbAkiMko6fWH.png)