remyxai/flowedit-flux-modular

FlowEdit for FLUX — training-free, inversion-free image editing (Modular Diffusers custom block)

Text-based image editing on off-the-shelf FLUX.1-dev — no inversion, no fine-tuning — as a Modular Diffusers custom block. Implements FlowEdit (arXiv:2412.08629): instead of inverting the source image, it builds an ODE that transports it directly from the source prompt to the target prompt — more faithful to the original (structure preserved) and faster than inversion-based editing.

— edit your own photo by changing the prompt.

_{One source image, two edits (matched source prompt → target prompt) — the subject changes, the pose / composition / background are preserved. Training-free, no inversion.}

Usage

import torch
from diffusers import ModularPipeline

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

edited = pipe(
    image="cat.png",
    source_prompt="a cat",       # describes the source image
    prompt="a dog",              # the target edit
    T_steps=28, src_guidance_scale=1.5, tar_guidance_scale=5.5, n_max=24, n_min=0,
).images[0]
edited.save("edited.png")

Tip: the source_prompt should describe what's actually in the image; the prompt is the edit. A good source_prompt is what makes the edit faithful. No extra weights or models — it's fully training-free.

How it works

FlowEdit is inversion-free. At each step it draws noise to form a source point zt_src and a coupled target point zt_tar, computes the model's velocity for each under the source and target prompts, and integrates the guided velocity difference Vt_tar − Vt_src into a running edit latent — over a step window (n_max … n_min), averaged over n_avg noise draws. Because it never inverts the image, the edit stays anchored to the original structure. A self-contained custom denoise loop; the base weights are untouched.

Key parameters

arg	default	meaning
image	—	source image (path / PIL / numpy RGB)
source_prompt	—	description of the source image
prompt	—	the target edit
tar_guidance_scale	5.5	edit strength (↑ = stronger edit)
src_guidance_scale	1.5	source guidance
T_steps	28	total steps
n_max / n_min	24 / 0	step window where the edit ODE is applied (raise n_min for an SDEdit-style tail)
n_avg	1	velocity-difference averaging (↑ = smoother, slower)

Attribution & AI assistance

Training-free reimplementation for Modular Diffusers of FlowEdit (MIT, fallenshock/FlowEdit). The Modular-Diffusers adaptation was authored with AI assistance (Claude) and validated by the Remyx AI team; method credit to the FlowEdit authors. Uses FLUX.1-dev under its non-commercial license.

Citation

@misc{kulikov2024flowedit,
  title={FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models},
  author={Kulikov, Vladimir and Kleiner, Matan and Huberman-Spiegelglas, Inbar and Michaeli, Tomer},
  year={2024},
  eprint={2412.08629}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2412.08629}
}