---
library_name: diffusers
pipeline_tag: image-to-image
tags:
  - polarization
  - image-to-image
  - diffusion
  - controlnet
  - lora
---

# GenPolar

Official model weights for **GenPolar: Stokes-Informed Diffusion for Robust Linear Polarization Estimation**.

GenPolar estimates the linear Stokes components from an intensity/RGB observation and derives the degree and angle of linear polarization analytically.

## Released files

| File | Purpose | Required for final inference |
| --- | --- | --- |
| `genpolar_one_step.pth` | Distilled one-step UNet and ControlNet | Yes |
| `genpolar_vae_encoder_lora/` | VAE encoder LoRA adapter used by the one-step model | Yes |
| `genpolar_stage1_teacher.pth` | Stage-one diffusion teacher used for distillation and stage-one reproduction | No |

The `.pth` files are inference-focused exports. They contain only `unet_state_dict` and `controlnet_state_dict`, plus small format metadata. Optimizer states, training counters, the fake score model, training arguments, and duplicate state dictionaries are intentionally excluded.

All released tensors retain their original FP32 values.

## Base model

The code initializes model components from `runwayml/stable-diffusion-v1-5` before loading the GenPolar state dictionaries.

## One-step inference

Place the downloaded files under `release_weights/`, then run:

```bash
python inference.py \
  --ckpt_path ./release_weights/genpolar_one_step.pth \
  --vae_lora_dir ./release_weights/genpolar_vae_encoder_lora \
  --input_folder ./data/DfP \
  --results_folder ./results \
  --fp16
```

The input directory format and full environment setup are documented in the GenPolar source repository.

## Integrity

SHA-256 hashes are provided in `SHA256SUMS`.