🏃SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation

     

How to Inference

1. How to use SanaSprintPipeline with 🧨diffusers

Note

Upgrade your diffusers to use SanaSprintPipeline:

pip install git+https://github.com/huggingface/diffusers

# test sana sprint
from diffusers import SanaSprintPipeline
import torch

pipeline = SanaSprintPipeline.from_pretrained(
    "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers",
    torch_dtype=torch.bfloat16
)
# Use DC-AE-Lite for faster speed.
# from diffusers import AutoencoderDC
# vae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers")
# pipeline.vae = vae
pipeline.to("cuda:0")

prompt = "a tiny astronaut hatching from an egg on the moon"

image = pipeline(prompt=prompt, num_inference_steps=2).images[0]
image.save("test_out.png")

# if you want to compile the vae. You need to upgrade to torch>=2.6.0
# DCAE1.1: 1287MB/0.12s; DCAE1.1Lite:11299MB/0.06s; DCAE1.1Lite compile: 10385MB/0.03s
import torch
from diffusers import AutoencoderDC

torch._dynamo.config.force_parameter_static_shapes = False
torch._dynamo.config.dynamic_shapes = True
torch._dynamo.config.recompile_limit = 16

vae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-lite-f32c32-sana-1.1-diffusers").to('cuda')
vae.decode = torch.compile(vae.decode, dynamic=True)

2. How to use SanaSprintPipeline in this repo

import torch
from app.sana_sprint_pipeline import SanaSprintPipeline
from torchvision.utils import save_image

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
generator = torch.Generator(device=device).manual_seed(42)

sana = SanaSprintPipeline("configs/sana_sprint_config/1024ms/SanaSprint_1600M_1024px_allqknorm_bf16_scm_ladd.yaml")
sana.from_pretrained("hf://Efficient-Large-Model/Sana_Sprint_1.6B_1024px/checkpoints/Sana_Sprint_1.6B_1024px.pth")

prompt = "a tiny astronaut hatching from an egg on the moon",

image = sana(
    prompt=prompt,
    height=1024,
    width=1024,
    guidance_scale=4.5,
    num_inference_steps=2,
    generator=generator,
)
save_image(image, 'sana_sprint.png', nrow=1, normalize=True, value_range=(-1, 1))

How to Train

bash train_scripts/train_scm_ladd.sh \
      configs/sana_sprint_config/1024ms/SanaSprint_1600M_1024px_allqknorm_bf16_scm_ladd.yaml
      --data.data_dir="[data/toy_data]" \
      --data.type=SanaWebDatasetMS \
      --model.multi_scale=true \
      --data.load_vae_feat=true \
      --train.train_batch_size=2

Convert pth to diffusers safetensor

python scripts/convert_scripts/convert_sana_to_diffusers.py \
      --orig_ckpt_path Efficient-Large-Model/Sana_Sprint_1.6B_1024px/checkpoints/Sana_Sprint_1.6B_1024px.pth \
      --model_type SanaSprint_1600M_P1_D20 \
      --scheduler_type scm \
      --dtype bf16 \
      --dump_path output/Sana_Sprint_1.6B_1024px_diffusers \
      --save_full_pipeline

Performance

Methods (1024x1024)	Inference Steps	Throughput (samples/s)	Latency (s)	Params (B)	FID 👇	CLIP 👆	GenEval 👆
Sana-Sprint_0.6B	2	6.46	0.25	0.6	6.54	28.40	0.76
Sana-Sprint-1.6B	2	5.68	0.24	1.6	6.50	28.45	0.77

---

Citation

@misc{chen2025sanasprint,
      title={SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation},
      author={Junsong Chen and Shuchen Xue and Yuyang Zhao and Jincheng Yu and Sayak Paul and Junyu Chen and Han Cai and Enze Xie and Song Han},
      year={2025},
      eprint={2503.09641},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.09641},
}