---
title: PG-MAP Demo
emoji: 🎨
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.19.0
python_version: "3.12"
app_file: app.py
pinned: true
license: mit
short_description: PG-MAP inference-time alignment (NeurIPS 2026 submission)
---

# PG-MAP Demo · NeurIPS 2026 (under review)

Interactive demo for **PG-MAP** (Preference-Guided Adaptive MAP) — a training-free framework that re-optimizes the conditioning $c$ and the latent $z_t$ at every denoising step. Supports SD 1.5, SDXL, and SD3.5-medium (UG-FM) backbones.

- 🔗 **Paper**: [arXiv:2606.22958](https://arxiv.org/abs/2606.22958) (under review at NeurIPS 2026) · [code](https://github.com/sophialanlan/PG-MAP)
- 🤗 **Custom diffusers pipelines**: [pg-map-sd15](https://huggingface.co/sophialan/pg-map-sd15) · [pg-map-sdxl](https://huggingface.co/sophialan/pg-map-sdxl) · [pg-map-sd3](https://huggingface.co/sophialan/pg-map-sd3)
- 📦 **PyPI**: `pip install pgmap-align`

## Hardware

This Space runs on the free **CPU-basic** tier, so the hosted demo cannot generate images as-is. To generate, duplicate the Space (or open **Settings → Hardware**) and select a GPU runtime: SDXL and SD3.5 need A10G small (24 GB VRAM) or larger; SD 1.5 fits on T4.

## Local development

```bash
git clone https://huggingface.co/spaces/sophialan/pg-map-demo
cd pg-map-demo
pip install -r requirements.txt
python app.py
```

## Citation

```bibtex
@misc{sun2026pgmap,
  title={{PG-MAP}: Joint {MAP} Optimization for Inference-Time Alignment of Diffusion and Flow-Matching Models},
  author={Sun, Ruolan and Polak, Pawel},
  year={2026},
  eprint={2606.22958},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  note={Under review at NeurIPS 2026},
  url={https://arxiv.org/abs/2606.22958}
}
```