Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods
Paper • 2508.02363 • Published • 1
How to use maryanlup/OT-RF with Diffusers:
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("maryanlup/OT-RF", dtype=torch.bfloat16, device_map="cuda")
prompt = "Turn this cat into a dog"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png")
image = pipe(image=input_image, prompt=prompt).images[0]Paper · arXiv · Code · Demo · Video
Zero-shot, OT-guided inversion and editing for rectified flow models, built on 🤗 Diffusers. This repo hosts only the pipeline code; weights come from FLUX.1-dev (gated — accept its license first).
import os, sys, torch
from diffusers import FluxPipeline
from diffusers.utils import load_image
from huggingface_hub import hf_hub_download
torch.manual_seed(42)
base_pipe = FluxPipeline.from_pretrained(
"black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16,
custom_pipeline="pipeline_flux_rf_inversion").to("cuda")
sys.modules["pipeline_flux_rf_inversion"] = sys.modules[type(base_pipe).__module__]
sys.path.append(os.path.dirname(hf_hub_download("maryanlup/OT-RF", "improved_optimal_transport_rf_inversion.py")))
from improved_optimal_transport_rf_inversion import OptimalTransportRFInversionPipeline
ot_pipe = OptimalTransportRFInversionPipeline.from_pipe(base_pipe)
image = load_image("https://raw.githubusercontent.com/marianlupascu/OT-RF/main/assets/face_input.jpg").resize((1024, 1024))
inverted_latents, image_latents, latent_image_ids = ot_pipe.invert(image=image, num_inversion_steps=28)
edited_image = ot_pipe(
prompt="a person wearing glasses",
inverted_latents=inverted_latents, image_latents=image_latents, latent_image_ids=latent_image_ids,
start_timestep=0, stop_timestep=7/28, num_inference_steps=28, eta=0.9,
use_ot_guidance=True, transport_strength=0.3, ot_guidance_phase=0.3,
respect_prompt_guidance=False,
).images[0]
edited_image.save("edited_ot.png")
See the GitHub README for the full inversion + editing example.
@InProceedings{Lupascu_2026_WACV,
author = {Lupa\c{s}cu, Marian and Stupariu, Mihai Sorin},
title = {Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2026},
pages = {6764--6774}
}