Flower

This project contains two experiments based on FLOWER:

1. A 2D toy experiment with a conditional GMM prior (gmm.ipynb)
2. A face image super-resolution experiment using a pretrained model (faces.ipynb)

The goal is to explain and visualize the 3-step FLOWER scheme:

1. Flow-consistent destination estimation
2. Refining this destination using the measurements
3. Updating the trajectory over time

The results are then compared with simple baselines.

---

Repository Structure

• gmm.ipynb: 2D GMM experiment, flow matching training, trajectory visualization
• faces.ipynb: face super-resolution experiment using pretrained FLOWER weights

Scripts

• scripts/create_gmm_data.py: generation of the toy GMM prior and posterior
• scripts/train_gmm_flow_matching.py: training of the flow matching model on posterior samples
• scripts/gmm_flow_model.py: time-conditioned MLP architecture + Euler sampling
• scripts/flower_steps.py: factorized implementation of the FLOWER steps (GMM and image inverse problems)
• scripts/flower_plotting.py: plotting utilities for the GMM experiment
• scripts/faces_pipeline.py: utilities for images, seeds, PSNR computation, and local super-resolution

Other folders

• results_examples/: example images already generated to illustrate outputs
• mehrsapo_Flower/: FLOWER source code used by the faces.ipynb experiment

---

Experiment 1: 2D GMM (gmm.ipynb)

1. Generates a 2D GMM prior and its posterior conditioned on a noisy linear observation
2. Trains a flow matching model on posterior samples
3. Applies the three FLOWER steps and saves step-by-step visualizations

---

Experiment 2: Face Super-Resolution (faces.ipynb)

1. Loads a pretrained FLOWER model trained on CelebA (weights are downloaded if missing)
2. Takes a custom face image and generates a noisy low-resolution observation
3. Reconstructs the high-resolution image using the FLOWER steps
4. Compares results with simple baselines (adjoint, bicubic) using PSNR