CheX-ray Super-Resolution (MUNIT-LQ → DDPM-HQ)

This repo contains code for our CheX-ray14 super-resolution experiments:

• MUNIT-LQ: low-quality (LQ) inputs generated from the CheX-ray14 test set (external MUNIT pipeline).
• DDPM-HQ: a diffusion-based super-resolution model (CheffSR/SR3-style) trained / fine-tuned on HQ–LQ pairs to map LQ → HQ.
• DDPM-HQ outputs: final super-res outputs produced by the DDPM-HQ model.

Most relevant entrypoints:

• Inference: cheff/sr/sampler.py, inference_batch_ddp.py
• Training (DDP): Train/train_ddpm_e2e.py, Train/train_sr_ddpm.py

---

Environment

We use the conda environment spec at cascade7.yaml:

conda env create -f cascade7.yaml
conda activate cascade7

---

Where the data & outputs live (server paths)

Outputs used in our CheX-ray14 runs

• Diffusion-based baseline (bicubic SR): /raid1/data_transfer/data/CheX-ray14/test_chexraybase_sr_bicubic/
• DDPM-HQ outputs (DDPM on LQ inputs) (per-GPU folder): /raid1/data_transfer/data/CheX-ray14/test_lr_ddpm_2/folder_x/
• MUNIT-LQ outputs: /raid2/data_transfer/data/CheX-ray14/test_lr_munit/all_images/all_images/
• DDPM-HQ outputs: /raid1/data_transfer/data/CheX-ray14/test_munit_sr_ddpm/

Curated qualitative bundle (100 matched images)

• Folder: /raid/collected_100_4way/
  • diffusion_baseline/
  • ddpm_lq/
  • munit_lq/
  • final_with_masks/ (DDPM-HQ outputs)
  • low_res_latest_160k/ (a 100-image low-res subset)
• Zip (4-way only): /raid/collected_100_4way.zip
• Zip (4-way + low-res): /raid/collected_100_4way_with_lowres.zip

---

MUNIT-LQ (low-quality generation)

Reference output path (already generated on servers):

• /raid2/data_transfer/data/CheX-ray14/test_lr_munit/all_images/all_images/

Notes:

• Filenames are like 00000013_016.png (no super_resolved_ prefix).
• These images are used as LQ inputs to the DDPM-HQ SR model.

---

DDPM-HQ training (super-resolution diffusion, HQ–LQ)

We train / fine-tune a CheffSR/SR3 diffusion model on HQ–LQ pairs to map LQ → HQ.

Baseline SR3 training (bicubic-conditioned) on MaCheX-style data

Entrypoints:

• Train/train_ddpm_e2e.py
• Train/scripts/04_train_chex_ddpm_ddp.py (same training loop style)

Run (important: diffusion.* imports live under Train/, so run from Train/):

cd /raid1/data_transfer/data/chexray-diffusion/Train
PYTHONPATH=. python train_ddpm_e2e.py

Update paths inside BASE_CONFIG:

• data_root: MaCheX root (expects subfolders each containing an index.json)
• resume_checkpoint: optional checkpoint to resume from
• log_dir: where checkpoints are saved

DDPM-HQ fine-tuning on our “low_res_latest” folder

Entrypoint:

• Train/train_sr_ddpm.py

By default it expects:

• data_root/low_res_latest/ (PNG/JPG images)

In our setup, the latest low-res pool is here:

• /raid2/data_transfer/low_res_latest_160k/

If you want to use it without changing code, create a symlink named low_res_latest next to it:

ln -s /raid2/data_transfer/low_res_latest_160k /raid2/data_transfer/low_res_latest

Run:

cd /raid1/data_transfer/data/chexray-diffusion/Train
PYTHONPATH=. python train_sr_ddpm.py

Update paths inside BASE_CONFIG in Train/train_sr_ddpm.py:

• data_root: parent directory that contains low_res_latest/
• resume_checkpoint: base checkpoint to fine-tune from
• log_dir: output checkpoint folder

---

Inference (DDPM-HQ)

Multi-GPU batch inference on a folder (recommended)

Entrypoint:

• inference_batch_ddp.py

Edit the hardcoded variables at the bottom of the file:

• sr_path: checkpoint to load
• input_folder: folder containing input images (.png / .jpg)
• output_folder: where to write results
• batch_size: per-GPU batch size

Run:

cd /raid1/data_transfer/data/chexray-diffusion
python inference_batch_ddp.py

Outputs are saved as super_resolved_<original_filename> in output_folder.

Example (CheX-ray14 MUNIT-LQ → DDPM-HQ outputs):

• input_folder: /raid2/data_transfer/data/CheX-ray14/test_lr_munit/all_images/all_images/
• output_folder: /raid1/data_transfer/data/CheX-ray14/test_lr_ddpm_2/folder_1/ (or any folder_x/)

Programmatic inference helper (single GPU)

The core API is implemented in:

• cheff/sr/sampler.py (CheffSRModel.sample_path(...), CheffSRModel.sample_directory(...))

---

Notes / gotchas

• Training script imports: Train/train_*.py imports diffusion.* from Train/diffusion/, so either:
  • run from chexray-diffusion/Train with PYTHONPATH=., or
  • set PYTHONPATH=/raid1/data_transfer/data/chexray-diffusion/Train.
• inference.py is a minimal snippet and expects an in-memory imgs tensor; for folder inference prefer inference_batch_ddp.py or CheffSRModel.sample_directory(...).