SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control

Paper | Project Page

We introduce SliderEdit, a framework for continuous image editing with fine-grained, interpretable instruction control. Given a multi-part edit instruction, SliderEdit disentangles the individual instructions and exposes each as a globally trained slider, allowing smooth adjustment of its strength.

⚙️Setup

1. Clone the repository

   git clone git@github.com:ArmanZarei/SliderEdit.git
   cd SliderEdit

2. Create and activate the conda environment

   conda env create -f environment.yml
   conda activate slideredit
   pip install -e .

🚀Quick Start

First, load the SliderEdit pipeline:

from slideredit.pipelines import SliderEditFluxKontextPipeline

pipe = SliderEditFluxKontextPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Kontext-dev",
    torch_dtype=torch.bfloat16
).to("cuda")

GSTLoRA

GSTLoRA is designed for single-instruction editing scenarios or when a single slider is sufficient to control the overall edit intensity.

pipe.load_gstlora("PATH_TO_CKPT")

output_image = pipe(
    image=IMAGE_TO_BE_EDITED,
    prompt="EDIT PROMPT",
    generator=torch.Generator().manual_seed(SEED)
    slider_alpha=STRENGH_VALUE,
).images[0]

The parameter slider_alpha controls edit strength. Negative values increase intensity, while positive values suppress the effect. We recommend initially sweeping values in the range [-1, 1].

STLoRA

STLoRA is designed for multi-instruction editing prompts, providing independent control over the strength of each individual instruction.

pipe.load_stlora("PATH_TO_CKPT")

output_image = pipe(
    image=IMAGE_TO_BE_EDITED,
    prompt="Edit_Instruction_1 and Edit_Instruct_2 and ...",
    generator=torch.Generator().manual_seed(SEED)
    subprompts_list=["Edit_Instruction_1", "Edit_Instruct_2", ...],
    slider_alpha_list=[Strength_Value_1, Strength_Value_2, ...],
).images[0]

Parameters:

• subprompts_list: Individual instructions (sub-prompts) from the original edit prompt
• slider_alpha_list: Corresponding intensity values for each instruction (we recommend initially sweeping values in the range: [-1, 1])

See sample_inference.ipynb for complete inference examples.

🔬Training SliderEdit

GSTLoRA

The training script for GSTLoRA is available at train_gstlora_flux_kontext.py. An example configuration using an open-source image editing dataset is provided in train_gstlora_flux_kontext.yaml.

To launch training:

python training/train_gstlora_flux_kontext.py --config="training/configs/train_gstlora_flux_kontext.yaml"

View training progress and slider visualizations in this W&B report.

STLoRA

Training STLoRA with our proposed PPS loss requires multi-instruction edit prompts. (If you instead wish to use the SPPS loss, the same dataset as in the GSTLoRA setting can be reused)

Below, we provide a simple small-scale example illustrating how to construct such a dataset and perform training. In this example, we focus on human face editing by manually combining a set of predefined single-instruction edits to form multi-instruction prompts.

For more details, please refer to the training script train_stlora_flux_kontext.py and the example configuration file train_stlora_pps_flux_kontext.yaml.

First, download the sample human faces dataset:

mkdir -p datasets
cd datasets
gdown 183-Jubsu2rFiQmgpBYjFpAxGSwqn2XOM
unzip slideredit_faces_dataset.zip
cd ..

Then, launch training:

python training/train_stlora_flux_kontext.py --config="training/configs/train_stlora_pps_flux_kontext.yaml"

View training progress and slider visualizations in this W&B report.

Inference

Below are example checkpoints from models trained using the above configurations:

mkdir -p checkpoints
gdown 1YHrHhSeKovEPGpFgFbv0iPL67YRgg6rG -O checkpoints/ # GSTLoRA iter500
gdown 1PdORTgzFzfGGbNAoPQb0T5su3t82xErY -O checkpoints/ # STLoRA iter1200

Example results:

• GSTLoRA

  
  "Make her hair curly"

• STLoRA

  
  "make the person fat and make the person laugh"

See sample_inference.ipynb for more details.

📊Evaluation

Simply instantiate any of the VLMEvaluator or FeatureDistanceEvaluator classes from the evaluation module and use them to compute the corresponding scores or distances. For more details and usage examples, please refer to evaluation/example_evaluation.ipynb.

BibTeX

@article{zarei2025slideredit,
  title={SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control},
  author={Zarei, Arman and Basu, Samyadeep and Pournemat, Mobina and Nag, Sayan and Rossi, Ryan and Feizi, Soheil},
  journal={arXiv preprint arXiv:2511.09715},
  year={2025}
}