You can deploy this model and start earning money today!

sophialan/
pg-map-sd3

Name: pg-map-sd3
Availability: InStock

inference-time-alignment

preference-optimization

Views

No views yet

PG-MAP / UG-FM for Stable Diffusion 3.5-medium

Custom diffusers pipeline for UG-FM — the flow-matching reduction of PG-MAP on SD3.5-medium. Defaults to the paper's headline configuration (data-side gate, $K_{UG}=4$, $\eta_z=0.1$, full backprop through the velocity prediction) which delivers 91.9% PickScore / 75.7% HPS win-rates against the static rectified-flow baseline on PartiPrompts ($n=1632$, seed 123).

Preprint — arXiv:2606.22958; see github.com/sophialanlan/PG-MAP for the paper, full configs, and reproduction scripts.

Install

bash

pip install pgmap-align
# or
pip install git+https://github.com/sophialanlan/PG-MAP

You also need to accept the Stability AI Community License for the SD3.5 weights on huggingface.co/stabilityai/stable-diffusion-3.5-medium before the first load.

Usage

python

from diffusers import DiffusionPipeline
from pgmap import FrozenRewardModel
import torch

pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-medium",
    custom_pipeline="sophialan/pg-map-sd3",
    torch_dtype=torch.float16,
).to("cuda")

reward = FrozenRewardModel("pickscore", device="cuda")

# UG-FM (default): 91.9% PickScore configuration
image = pipe(
    "a phoenix rising from ashes, vivid orange and red feathers",
    reward_model=reward,
).images[0]

For the full PG-MAP-FM (joint c + z_t with flow consistency + Gaussian priors + reward), pass pg_map_config with optimize_c=True:

python

from pgmap import sdxl_defaults
from dataclasses import replace

cfg = sdxl_defaults()                                     # starting point
cfg = replace(cfg, optimize_c=True, optimize_z=True)
image = pipe("a phoenix rising from ashes", pg_map_config=cfg).images[0]

Why UG-FM is the right default for flow matching

Per paper §3.2, on SD3.5 the optimal active set collapses to ${z_t}$ alone at data-side steps for two transport-specific reasons:

Conditioning capacity. SD3.5's concatenated CLIP-L / CLIP-G / T5-XXL representation has ~1.4 M optimisable parameters, so a unit-normalised c-gradient is spread too thinly to move any single direction.
Local Euler amplification. A noise-side perturbation traverses ~25 factors of $I + \Delta t_j,\partial_z v_\theta$ and grows 5–50×, while a data-side perturbation has only 1–3 remaining factors and stays bounded (sub-pixel mean RMSE $0.61/255$).

Paper headline (SD3.5-medium, PartiPrompts $n=1632$, seed 123)

Method	PickScore	HPS	Aesthetic	CLIP
Static baseline	50.0%	50.0%	50.0%	50.0%
FlowChef (always-on, K=1)	82.4%	68.1%	49.7%	53.9%
FlowChef (gating-matched)	75.0%	62.5%	46.9%	52.9%
UG-FM (Ours)	91.9%	75.7%	51.7%	54.2%

Win-rate vs. same-seed static baseline. The 16.9 pp PickScore gap between UG-FM and gating-matched FlowChef isolates the full backprop through $v_\theta$ axis — FlowChef's gradient skipping (with torch.no_grad(): v = v_theta(...)) discards the Jacobian factor $I - (1-t),\partial_z v_\theta$ which is load-bearing.

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},
  url={https://arxiv.org/abs/2606.22958}
}

License

MIT (see LICENSE). SD3.5 weights are under the Stability AI Community License.