TextFlow

English | 简体中文 | Paper

Code for Towards Training-Free Scene Text Editing. TextFlow combines Flow Manifold Steering (FMS) and Attention Boost (AttnBoost) for training-free scene text editing, with FLUX-Kontext and LongCat backends.

Installation

Requires Linux, Python 3.10/3.11, an NVIDIA GPU and Git.

python3 -m venv .venv
source .venv/bin/activate
pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

Choose a CUDA build compatible with your driver. Keep the pinned diffusers version in requirements.txt.

Models and data

Download the models separately into:

models/FLUX.1-Kontext-dev/
models/LongCat-Image-Edit/

FLUX accepts --model_path; LongCat uses the fixed path above. Prepare a JSON list:

[
  {"image_path": "./data/images/sample.png", "text_s": "OPEN", "text_t": "CLOSED"}
]

text_s is the original text and text_t is the replacement. Paths are relative to the repository root. Optional unique id fields distinguish repeated records.

Usage

Run from the repository root.

Single GPU

# FLUX-Kontext
CUDA_VISIBLE_DEVICES=0 python infer_flux_text.py \
  --json_path ./data/dataset.json --exp_dir ./outputs/flux \
  --scheduler overshoot --use_att --use_flowedit

# LongCat
CUDA_VISIBLE_DEVICES=0 python infer_longcat_text.py \
  --json_path ./data/dataset.json --exp_dir ./outputs/longcat \
  --scheduler overshoot --use_att --use_flowedit

Multiple GPUs

CUDA_VISIBLE_DEVICES=0,1 accelerate launch --multi_gpu --num_processes 2 \
  infer_flux_text_multi_gpu.py --json_path ./data/dataset.json \
  --exp_dir ./outputs/flux_multi --scheduler overshoot --use_att --use_flowedit

For LongCat, replace the entry with infer_longcat_text_multi_gpu.py. Each GPU loads a full model.

• --use_flowedit enables FMS; --use_att enables AttnBoost. Omit the former to disable FMS.
• To use Euler without attention, pass --scheduler euler --no-use_att.
• Results go under --exp_dir. Resume with the same settings and dataset; use a new directory when changing settings, code or models.

Entry points

Entry	Task
infer_flux_text.py	FLUX text replacement
infer_flux_text_multi_gpu.py	FLUX multi-GPU text replacement
infer_flux_color_multi_gpu.py	FLUX multi-GPU color editing (yellow)
infer_longcat_text.py	LongCat text replacement with optional FMS
infer_longcat_text_multi_gpu.py	LongCat multi-GPU text replacement
infer_longcat_attention.py	LongCat AttnBoost without FMS
infer_flux_baseline.py	Native FLUX-Kontext baseline; paths set in the script

Tests

CPU regression tests require only Python and Pillow:

python -B -m unittest discover -s tests -v
python -B -m unittest discover -s textflow_flux -p test_pipeline_safety.py -v
python -B -m unittest discover -s textflow_longcat -p test_pipeline_safety.py -v

CI · Changelog

License

Apache-2.0. Third-party attribution: NOTICE and diffusers license. Model weights and datasets are subject to their own terms.