Apppublic

azam897/Diffusion-Model

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

App README

DDPM: Denoising Diffusion Probabilistic Model

Trained on CelebA-HQ 256×256 · AI4009 Generative AI · Spring 2026 · NUCES

What this app does

Tab	Description
Generate from Noise	Runs the full reverse diffusion process starting from random Gaussian noise and produces a 256×256 face image. Supports both DDPM (full quality, 300 steps) and DDIM (fast, configurable steps).
Reconstruct Image	Upload any face image, partially corrupt it with forward diffusion, then reconstruct it using the model.

Model Details

Property	Value
Architecture	U-Net with Residual Blocks & Self-Attention
Channel progression	64 → 128 → 256
Residual blocks per level	2
Attention resolution	32×32
Timesteps	300
Noise schedule	Linear (β₀=0.0001 → β_T=0.02)
Dataset	CelebA-HQ 256×256
Training epochs	40
Best loss	0.0199
Optimizer	AdamW + CosineAnnealingLR
Loss function	MSE on predicted noise

Model Weights

Weights are loaded automatically at startup from 👉 `shahem1/AI4009-DDPM-Model` (best_model.pt)

Quick Usage (Python)

python

from huggingface_hub import hf_hub_download
import torch

ckpt = hf_hub_download("shahem1/AI4009-DDPM-Model", "best_model.pt")
checkpoint = torch.load(ckpt, map_location="cpu")

# Instantiate UNet with the same config, then:
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()