This repository contains the *official codebase for Kontinuous Kontext, a model for fine-grained strength control for generalized instruction-based image editing.

Rishubh Parihar, Or Patashnik, Daniil Ostashev, R. Venkatesh Babu, Daniel Cohen-Or, Kuan-Chieh (Jackson) Wang

Snap Research · Tel Aviv University · IISc Bangalore

Instruction-based image editing offers a powerful and intuitive way to manipulate images through natural language. Yet, relying solely on text instructions limits fine-grained control over the extent of edits. We introduce Kontinuous Kontext, an instruction-driven editing model that provides a new dimension of control over edit strength, enabling users to adjust edits gradually from no change to a fully realized result in a smooth and continuous manner. Kontinuous Kontext extends a state-of-the-art image editing model to accept an additional input, a scalar edit strength which is then paired with the edit instruction, enabling explicit control over the extent of the edit. To inject this scalar information, we train a lightweight projector network that maps the input scalar and the edit instruction to coefficients in the model's modulation space. For training our model, we synthesize a diverse dataset of image-edit-instruction-strength quadruplets using existing generative models, followed by a filtering stage to ensure quality and consistency. Kontinuous Kontext provides a unified approach for fine-grained control over edit strength for instruction driven editing from subtle to strong across diverse operations such as stylization, attribute, material, background, and shape changes, without requiring attribute-specific training.

1. Clone the Repository

   git clone git@github.sc-corp.net:Snapchat/kontinuous_kontext.git
   cd kontinuous-kontext

2. Create the Conda Environment Set up the required environment using the provided file:

   conda env create -f kslider_environment.yml

We have a few extras queued up for release—this section will grow as they roll out.

• Training and inference code release
• Release of the complete dataset for model training
• Scripts for dataset generation pipeline

Follow these steps to run inference with a pretrained model.

1. Download Model Weights Download the model weights from here and copy them into the ./model_weights/ folder.

2. Run Inference Execute the infer.sh script. This will run the edits on the images from the ./assets/ folder using the prompts defined in the Python script.

   bash infer.sh

The infer.sh script contains the following command.

python3 test_ksliders.py \
    --pretrained_model_name_or_path=black-forest-labs/FLUX.1-Kontext-dev \
    --trained_models_path=./model_weights \
    --input_images_path=./assets \
    --n_edit_steps=6 \
    --images_save_path=output_images

Spin up the interactive Gradio playground to try your own examples without leaving your machine.

1. Install dependencies (same environment used for inference works here as well).
2. Download model weights into ./model_weights/ if you haven’t already.
3. Launch the Gradio UI:

   python gradio_demo.py \
       --pretrained_model_name_or_path=black-forest-labs/FLUX.1-Kontext-dev \
       --trained_models_path=./model_weights \
       --input_images_path=/path/to/your/images

4. Open the printed local URL (defaults to http://127.0.0.1:7860) and upload an image plus an edit instruction. Adjust the strength slider to explore the continuum of edits.

_{Screenshot placeholder — drop in your latest Gradio UI preview when ready.}

Tip: Use --share when launching if you want to generate a temporary public Gradio link for collaborators.

We build our model on Flux Kontext, an instruction-based image editing model. For training, you have to download the Flux Kontext model and the training dataset as discussed below.

• You can download a sample dataset required for training from here. We are working on the full dataset and will release it soon.

• The dataset is organized into two parts:

  • sampling_data: Contains the image edit sequences generated by our data generation pipeline. It contains the source and full edit images, their inversions and interpolations in a single image stack as shown below.

  

  • sample_data_scores_w_scores.json: Contains the metadata for the slider dataset, including the mage name, edit instruction, LPIPS scores between adjacent images, and the LPIPS scores between the input and the inversion images for filtering. A sample entry from the .json file is shown below

    {
      "image_name": "image_0.png",
      "extended_image_name": "image_0_appearance_change | Add_glowing_fireflies_to_the_palm_tree_fronds.png",
      "category": "appearance_change",
      "edit_instruction": "Add_glowing_fireflies_to_the_palm_tree_fronds",
      "lpips_kontext_edit": 0.05086366832256317,  # Distance between the original source and the full edit output of Kontext
      "lpips_edit_inversion": 0.3724202513694763, # Distance between the original full edit and its inversion 
      "lpips_inversion_edit": 0.35860997438430786, # Distance between the inversion of the source and the inversion full edit output of Kontext
      "lpips_sequence": [
          0.2703213691711426,
          0.16677162051200867,
          0.1193908229470253,
          0.03886914253234863,
          0.05533060431480408,
          0.08398035168647766,
          0.18131063878536224,
          0.3724202513694763
      ]
  },
  

• The dataset will be filtered in the slider_dataset.py by loading the sample_data_scores_w_scores.json file. You can also adjust the filtering criteria based on your requirements.

We trained our model on 8 x NVIDIA A100 (80G) GPU for 110,000 iterations. The training typically requires 3-4 days on this hardware.

You can start a training run by executing the train.sh script. The base FLUX.1-Kontext-dev model will be downloaded automatically by the script on the first run.

bash train.sh

Your training logs will be saved inside the ./logs/ folder.

Here is the full content of the training script, which includes all hyperparameters used for our model:

RUN_NAME="TRAINING_RUN"

accelerate launch train_model.py \
  --pretrained_model_name_or_path=black-forest-labs/FLUX.1-Kontext-dev \
  --output_dir=runs/$RUN_NAME \
  --mixed_precision="bf16" \
  --resolution=512 \
  --data_json_path="PATH for the json file for the dataset" \
  --image_dataset_path="Dataset path" \
  --kl_threshold=0.15 \
  --filter="kl-filter-simple" \
  --train_batch_size=1 \
  --guidance_scale=1 \
  --optimizer="adamw" \
  --use_8bit_adam \
  --learning_rate=2e-5 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=200 \
  --max_train_steps=110000 \
  --checkpointing_steps=2000 \
  --report_to="wandb" \
  --run_name=$RUN_NAME \
  --drop_text_prob=0.0 \
  --validation_image_path="./src_imgs" \
  --num_validation_images=1 \
  --slider_projector_out_dim=6144 \
  --slider_projector_n_layers=4 \
  --modulation_condn=True \
  --is_clip_input=True \
  --seed="0" 2>&1 | tee logs/${RUN_NAME}.log

@article{parihar2025kontinuouskontext,
  title     = {Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing},
  author    = {Parihar, Rishubh and Patashnik, Or and Ostashev, Daniil and Babu, R. Venkatesh and Cohen-Or, Daniel and Wang, Kuan-Chieh Jackson},
  journal   = {arXiv preprint arXiv:2510.08532},
  year      = {2025},
  url       = {https://arxiv.org/abs/2510.08532}
}