remyxai/story-diffusion-flux-modular

StoryDiffusion for FLUX — training-free consistent-character generation (Modular Diffusers custom block)

Generate a set of frames of the same character across different scenes — no fine-tuning — as a Modular Diffusers custom block. Implements StoryDiffusion's Consistent Self-Attention (arXiv:2405.01434) on FLUX: each frame also attends to the other frames, so the character stays consistent while the scenes vary. Optional comic-sheet layout. Training-free.

— make a consistent-character comic from your prompts.

_{One character (prompt) across Morning / Coffee / Afternoon / Evening scenes — consistent identity, distinct scenes. Training-free.}

Usage

import torch
from diffusers import ModularPipeline

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

out = pipe(
    character_prompt="a young woman with curly red hair and freckles, green jacket",
    scene_prompts=[
        "waking up in a sunlit bedroom #Morning",       # "#..." -> panel caption
        "drinking coffee in a cozy kitchen #Coffee",
        "walking through a city park #Afternoon",
        "reading in a warm cafe at night #Evening",     # "[NC] ..." -> a scene without the character
    ],
    comic_layout="grid", comic_cols=2,
    height=1024, width=1024,
).images
sheet, panels = out[0], out[1:]      # comic_layout -> images[0] is the composed sheet
sheet.save("comic.png")

How it works

The frames are generated together with a Consistent Self-Attention processor: each frame keeps its own tokens and additionally attends to a sampled fraction (share_ratio=0.3) of the other frames' tokens, but only after the first ~third of steps (share_start_frac=0.35) — so each scene's composition is set first, then the shared character is locked in. This keeps scenes distinct while the character stays consistent. The base weights are untouched (restored after). Scope: image consistency + comic compositor (no video; PhotoMaker real-face identity is a possible v2 that pairs with remyxai/pulid-flux-modular).

Key parameters

arg	default	meaning
character_prompt	—	the subject description shared across frames
scene_prompts	—	list; "#caption" sets a panel caption; "[NC]" prefix = scene without the character
share_ratio	0.3	fraction of cross-frame tokens shared (↑ = more consistent, less scene diversity)
share_start_frac	0.35	share only after this fraction of steps (scene set first)
comic_layout	None	"grid" composes a captioned comic sheet (images[0])
guidance_scale	3.5	FLUX guidance

Attribution & AI assistance

FLUX (MMDiT) port of StoryDiffusion's Consistent Self-Attention (HVision-NKU/StoryDiffusion, 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{zhou2024storydiffusion,
  title={StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation},
  author={Zhou, Yupeng and Zhou, Daquan and Cheng, Ming-Ming and Feng, Jiashi and Hou, Qibin},
  year={2024}, eprint={2405.01434}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2405.01434}
}