README.md

metadata

license: mit
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - diffusers
  - pmf
  - image-generation
  - class-conditional
  - imagenet
inference: true
widget:
  - output:
      url: pMF-H-32/demo.png

pMF-diffusers

Native diffusers implementation of Pixel Mean Flows (pMF). Each variant folder is self-contained:

• pipeline.py — PMFPipeline
• scheduler/scheduler_config.json — FlowMatchEulerDiscreteScheduler config
• transformer/transformer_pmf.py — PMFTransformer2DModel
• transformer/ — converted weights and config

Available checkpoints

Checkpoint	Path	Resolution	Recommended CFG (ω)	CFG interval	Noise scale
pMF-B/16	./pMF-B-16	256×256	7.5	[0.1, 0.8]	1.0
pMF-B/32	./pMF-B-32	512×512	6.5	[0.1, 0.7]	2.0
pMF-L/16	./pMF-L-16	256×256	7.0	[0.2, 0.7]	1.0
pMF-L/32	./pMF-L-32	512×512	7.5	[0.2, 0.6]	4.0
pMF-H/16	./pMF-H-16	256×256	7.0	[0.2, 0.6]	2.0
pMF-H/32	./pMF-H-32	512×512	5.5	[0.1, 0.6]	4.0

Inference

from pathlib import Path
from diffusers import DiffusionPipeline
import torch

model_dir = Path("./pMF-L-16")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    local_files_only=True,
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.float32,
).to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=7.0,
    guidance_interval_min=0.2,
    guidance_interval_max=0.7,
    noise_scale=1.0,
    generator=generator,
).images[0]
image.save("demo.png")

Load a variant subfolder (e.g. ./pMF-L-16), not the repo root.