PG-MAP Demo  ·  NeurIPS 2026 (under review)

Inference-time alignment for diffusion + flow-matching — re-optimize the conditioning $c$ and the latent $z_t$ at every denoising step under a trajectory-level Gibbs-MAP / proximal energy objective. No training required.

🔗 Code: github.com/sophialanlan/PG-MAP · Paper: arXiv:2606.22958 · HF Pipelines: sd15 · sdxl · sd3

Pick a backbone, write a prompt, hit Generate. Toggle PG-MAP off to compare against the static baseline at the same seed. Default hyperparameters match the paper table; the sliders expose the productive ranges.