非对称流模型

像素空间文本到图像模型 AsymFLUX.2-klein 基于 black-forest-labs/FLUX.2-klein-base-9B 进行微调，采用了论文中提出的 AsymFlow 方法：

非对称流模型
arXiv 2026
Hansheng Chen, Jan Ackermann, Minseo Kim, Gordon Wetzstein, Leonidas Guibas
斯坦福大学
项目页面 | arXiv | 代码 | AsymFLUX.2 klein 演示🤗

asymflow_teaser

使用方法

请先安装 LakonLab v0.2。

我们为 AsymFLUX.2 klein 提供了 Diffusers 风格的流水线。以下示例加载 FLUX.2 klein Base 9B 模型，附加 AsymFlow 适配器，并直接在像素空间生成图像。

import math
import torch
from lakonlab.models.architectures import OklabColorEncoder
from lakonlab.models.diffusions.schedulers import FlowAdapterScheduler
from lakonlab.pipelines.pipeline_pixelflux2_klein import PixelFlux2KleinPipeline

pipe = PixelFlux2KleinPipeline.from_pretrained(
    'black-forest-labs/FLUX.2-klein-base-9B',
    vae=OklabColorEncoder(
        use_affine_norm=True,
        mean=(0.56, 0.0, 0.01),
        std=0.16),
    scheduler=FlowAdapterScheduler(
        shift=17.0,
        use_dynamic_shifting=True,
        base_seq_len=1024 ** 2,
        max_seq_len=2048 ** 2,
        base_logshift=math.log(17.0),
        max_logshift=math.log(34.0),
        dynamic_shifting_type='sqrt',
        base_scheduler='UniPCMultistep'),
    torch_dtype=torch.bfloat16)
adapter_name = pipe.load_lakonlab_adapter(  # you may later call `pipe.set_adapters([adapter_name, ...])` to combine other adapters (e.g., style LoRAs)
    'Lakonik/AsymFLUX.2-klein-9B',
    target_module_name='transformer')
pipe = pipe.to('cuda')

# Text-to-image generation example
prompt = 'Restored color photo from the 1900s. A middle-aged man with cybernetic metal hands is sitting on an old wooden chair and reading the newspaper. The newspaper has the prominent headline "AsymFLOW RELEASED" in large bold font. Close-up shot focusing on the newspaper.'
neg_prompt = 'Low quality, worst quality, blurry, deformed, bad anatomy, unclear text'
out = pipe(
    prompt=prompt,
    negative_prompt=neg_prompt,
    width=960,
    height=1280,
    num_inference_steps=38,
    guidance_scale=4.0,
    generator=torch.Generator().manual_seed(42),
).images[0]
out.save('asymflux2_klein.png')

引用格式

@article{chen2026asymmetric,
  title={Asymmetric Flow Models},
  author={Hansheng Chen and Jan Ackermann and Minseo Kim and Gordon Wetzstein and Leonidas Guibas},
  journal={arXiv preprint arXiv:2605.12964},
  url={https://arxiv.org/abs/2605.12964},
  year={2026},
}

使用方法

我们为 AsymFLUX.2 klein 提供了 Diffusers 风格的流水线。以下示例加载 FLUX.2 klein Base 9B 模型，附加 AsymFlow 适配器，并直接在像素空间生成图像。

import math
import torch
from lakonlab.models.architectures import OklabColorEncoder
from lakonlab.models.diffusions.schedulers import FlowAdapterScheduler
from lakonlab.pipelines.pipeline_pixelflux2_klein import PixelFlux2KleinPipeline

pipe = PixelFlux2KleinPipeline.from_pretrained(
    'black-forest-labs/FLUX.2-klein-base-9B',
    vae=OklabColorEncoder(
        use_affine_norm=True,
        mean=(0.56, 0.0, 0.01),
        std=0.16),
    scheduler=FlowAdapterScheduler(
        shift=17.0,
        use_dynamic_shifting=True,
        base_seq_len=1024 ** 2,
        max_seq_len=2048 ** 2,
        base_logshift=math.log(17.0),
        max_logshift=math.log(34.0),
        dynamic_shifting_type='sqrt',
        base_scheduler='UniPCMultistep'),
    torch_dtype=torch.bfloat16)
adapter_name = pipe.load_lakonlab_adapter(  # you may later call `pipe.set_adapters([adapter_name, ...])` to combine other adapters (e.g., style LoRAs)
    'Lakonik/AsymFLUX.2-klein-9B',
    target_module_name='transformer')
pipe = pipe.to('cuda')

# Text-to-image generation example
prompt = 'Restored color photo from the 1900s. A middle-aged man with cybernetic metal hands is sitting on an old wooden chair and reading the newspaper. The newspaper has the prominent headline "AsymFLOW RELEASED" in large bold font. Close-up shot focusing on the newspaper.'
neg_prompt = 'Low quality, worst quality, blurry, deformed, bad anatomy, unclear text'
out = pipe(
    prompt=prompt,
    negative_prompt=neg_prompt,
    width=960,
    height=1280,
    num_inference_steps=38,
    guidance_scale=4.0,
    generator=torch.Generator().manual_seed(42),
).images[0]
out.save('asymflux2_klein.png')

@article{chen2026asymmetric, title={Asymmetric Flow Models}, author={Hansheng Chen and Jan Ackermann and Minseo Kim and Gordon Wetzstein and Leonidas Guibas}, journal={arXiv preprint arXiv:2605.12964}, url={https://arxiv.org/abs/2605.12964}, year={2026}, }