ddpm-celebahq-256-finetuned-100steps-20251204

Fine-tuned DDPM model based on google/ddpm-celebahq-256.

Training Details

• Base Model: google/ddpm-celebahq-256
• Training Scenario: representative_mix (20% clean + 80% corrupted CelebA-HQ)
• Corruptions: 4 representative types (gaussian_noise, defocus_blur, fog, jpeg_compression) at severity 3
• Training Steps: 100
• Final Loss: 0.036426

Files

• unet_model.pth: Fine-tuned UNet model weights
• metadata.json: Training metadata and configuration

Usage

Load this checkpoint using the CheckpointManager from the training repository.

from checkpoint_manager import CheckpointManager
from diffusers import DDPMPipeline

# Load base model
pipe = DDPMPipeline.from_pretrained("google/ddpm-celebahq-256")

# Load fine-tuned weights
checkpoint_manager = CheckpointManager(
    base_model_name="google/ddpm-celebahq-256",
    local_checkpoint_dir="./checkpoints_celebahq_representative_mix"
)
checkpoint_manager.load_checkpoint_local("ddpm-celebahq-256-finetuned-100steps-20251204", pipe.unet)