Priming-diffusers

Native diffusers checkpoints for Priming (Are We Overthinking One-Step Generation?). Code: JaeyeonKim01/priming. Each variant folder is self-contained:

• pipeline.py — PrimingPipeline
• scheduler/scheduling_priming.py — PrimingScheduler (one evaluation; the transformer predicts the clean image)
• transformer/transformer_priming.py — PrimingTransformer2DModel

Load with PyPI diffusers and trust_remote_code=True. No extra Priming package is required. Sampling is one forward pass in RGB at 256×256 with guidance_scale=1. Weights are the released post-training EMA. The paper reports ImageNet-256 FID 1.38 for Priming-B and 1.19 for Priming-L.

Available checkpoints

Checkpoint	Path	Resolution	Steps	Guidance
Priming-B	Priming-B/	256×256	1	1.0
Priming-L	Priming-L/	256×256	1	1.0

Demo

Class: golden retriever (ImageNet 207) — 256×256, 1 step, guidance_scale=1.0, torch_dtype=bfloat16, seed 1.

ImageNet class labels

Each variant keeps an English id2label map in model_index.json.

• pipe.id2label — id → English label
• pipe.labels — synonym → id
• pipe.get_label_ids("golden retriever")
• pipe(class_labels="golden retriever", ...) — string labels resolve automatically

height and width default to 256. Other sizes divisible by the patch size of 16 interpolate the learned positions and rebuild 2-D rotary frequencies.

Inference

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/Priming-diffusers",
    subfolder="Priming-B",  # or "Priming-L"
    custom_pipeline="pipeline.py",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

print(pipe.id2label[207])
image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=1.0,
    generator=torch.Generator(device="cuda").manual_seed(1),
).images[0]
image.save("demo.png")

Local checkout:

from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("Priming-diffusers/Priming-B")
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.bfloat16,
).to("cuda")

Citation

@misc{park2026are,
  title   = {Are We Overthinking One-Step Generation?},
  author  = {Park, Woobin and Kim, Jaeyeon and Du, Yilun and Chen, Sitan and Han, Bohyung},
  year    = {2026},
  eprint  = {2610.20331},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url     = {https://arxiv.org/abs/2610.20331}
}