remyxai/style-aligned-flux-modular

StyleAligned for FLUX — training-free style-consistent set generation (Modular Diffusers custom block)

Generate a set of images that share one coherent style — no fine-tuning — as a Modular Diffusers custom block. Implements StyleAligned (arXiv:2312.02133) on FLUX: the batch shares attention so every image adopts the style of the first (anchor) prompt while keeping its own subject. Training-free; no extra weights.

— pick a style + subjects and generate a matching set.

_{"a red panda / a fox / a rabbit, flat vector sticker art" — distinct subjects, one shared style. Training-free.}

Usage

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("remyxai/style-aligned-flux-modular", trust_remote_code=True)
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

images = pipe(
    prompts=[
        "a red panda, flat vector sticker art",   # item 0 = the style anchor
        "a fox, flat vector sticker art",
        "a rabbit, flat vector sticker art",
    ],
    height=1024, width=1024,
).images
for i, im in enumerate(images):
    im.save(f"styled_{i}.png")

Put the style in every prompt; item 0 anchors it. style_share=False gives independent generations.

How it works

The batch is generated together with a shared-attention processor: each image keeps its own query (so its subject stays distinct), AdaIN-aligns its keys to item-0's statistics, and additionally attends to item-0's K/V — but only in the last ~20% of layers (share_start_frac=0.8), where style lives. Early/mid layers stay independent, so subjects don't collapse. The base weights are untouched (restored after).

Key parameters

arg	default	meaning
prompts	—	list of prompts sharing one style (item 0 = anchor)
style_share	True	False = independent generations
share_start_frac	0.8	share only the last (1-frac) of layers (lower = stronger style, risks subject collapse)
guidance_scale	3.5	FLUX guidance
num_inference_steps	28	denoise steps

Attribution & AI assistance

FLUX (MMDiT) port of StyleAligned (google/style-aligned, Apache-2.0). Authored with AI assistance (Claude) and validated by the Remyx AI team. Uses FLUX.1-dev under its non-commercial license.

Citation

@misc{hertz2023stylealigned,
  title={Style Aligned Image Generation via Shared Attention},
  author={Hertz, Amir and Voynov, Andrey and Fruchter, Shlomi and Cohen-Or, Daniel},
  year={2023}, eprint={2312.02133}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2312.02133}
}