sunnycloudhust/CelebA-flow-matching

Flow Matching for CelebA

This model learns a velocity field that transports Gaussian noise into 64x64 CelebA face images using flow matching.

Architecture

• Custom U-Net with sinusoidal time embeddings
• Input image size: 64x64
• Input channels: 3
• Base channels: 64

Loading

import torch
from modeling import FlowMatchingModel

model = FlowMatchingModel.from_pretrained(".")
model.eval()

Training setup

This project uses a minimal Euler ODE sampler with flow matching and a velocity target:

x_t = (1 - t) * x0 + t * x1
target_v = x1 - x0

The checkpoint in this repo is a raw PyTorch state dict generated by the training script.