README.md

metadata

base_model:
  - black-forest-labs/FLUX.1-dev
  - Qwen/Qwen-Image
license: apache-2.0
pipeline_tag: text-to-image
library_name: diffusers

ArcFlow

ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation
Zihan Yang¹, Shuyuan Tu¹, Licheng Zhang¹, Qi Dai², Yu-Gang Jiang¹, Zuxuan Wu¹
[¹Fudan University; ²Microsoft Research Asia]

Official Code Repository: pnotp/ArcFlow

Overview

ArcFlow is a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation.

Usage

Please first install the official code repository.

We provide diffusers pipelines for easy inference. The following code demonstrates how to sample images from the distilled FLUX.2 models.

4-NFE/2-NFE Arc-Qwen (Distilled from Qwen-Image-20B)

import torch
from diffusers import FlowMatchEulerDiscreteScheduler
from lakonlab.pipelines.arcqwen_pipeline import ArcQwenImagePipeline

pipe = ArcQwenImagePipeline.from_pretrained(
    'Qwen/Qwen-Image',
    torch_dtype=torch.bfloat16)
adapter_name = pipe.load_arcflow_adapter(  
    'ymyy307/ArcFlow',
    subfolder='arcflow-qwen-2steps',
    target_module_name='transformer')
pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(  # use fixed shift=3.2
    pipe.scheduler.config, shift=3.2, shift_terminal=None, use_dynamic_shifting=False)
pipe = pipe.to('cuda')

nfe = 4
# nfe = 2
out = pipe(
    prompt = 'A semi-realistic fantasy illustration featuring a split composition of two young men in profile, facing away from each other. On the left, a pale man with sharp features and slicked-back black hair wears a dark coat. On the right, a tan man with messy wavy hair wears a blue tunic. The ornate, 3D metallic gold title "Sultan\'s Game" overlays the bottom center. The background is divided into distinct sections: vibrant red abstract shapes in the upper half and deep teal textures in the lower half, creating a sharp color contrast. Painterly brushstrokes.',
    num_images_per_prompt=1,
    width=1024,
    height=1024,
    num_inference_steps=nfe,
    generator=torch.Generator(device="cuda").manual_seed(42),
    timestep_ratio=1.0,
).images[0]
out.save(f'arcqwen_{nfe}nfe.png')

4-NFE/2-NFE Arc-FLUX (Distilled from FLUX.1-dev)

import torch
from diffusers import FlowMatchEulerDiscreteScheduler
from lakonlab.pipelines.arcflux_pipeline import ArcFluxPipeline

pipe = ArcFluxPipeline.from_pretrained(
    'black-forest-labs/FLUX.1-dev',
    torch_dtype=torch.bfloat16)
adapter_name = pipe.load_arcflow_adapter(  # you may later call `pipe.set_adapters([adapter_name, ...])` to combine other adapters (e.g., style LoRAs)
    'ymyy307/ArcFlow',
    subfolder='arcflow-flux-2steps',
    target_module_name='transformer')
pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(  # use fixed shift=3.2
    pipe.scheduler.config, shift=3.2, shift_terminal=None, use_dynamic_shifting=False)
pipe = pipe.to('cuda')

nfe = 4
# nfe = 2
out = pipe(
    prompt = 'A portrait photo of a kangaroo wearing an orange hoodie and blue sunglasses standing in front of the Sydney Opera House holding a sign on the chest that says "Welcome Friends"',
    num_images_per_prompt=1,
    width=1024,
    height=1024,
    num_inference_steps=nfe,
    generator=torch.Generator(device="cuda").manual_seed(42),
    timestep_ratio=1.0,
).images[0]
out.save(f'arcflux_{nfe}nfe.png')

Citation

@misc{yang2026arcflowunleashing2steptexttoimage,
      title={ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation}, 
      author={Zihan Yang and Shuyuan Tu and Licheng Zhang and Qi Dai and Yu-Gang Jiang and Zuxuan Wu},
      year={2026},
      eprint={2602.09014},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.09014}, 
}