1. Overview

The Image Editing Pipeline generates corresponding edited images based on existing images and editing instructions. The usage is very simple: input image files and editing instructions (text prompts), and output edited images.

This version uses local GPU models for image editing, supporting local deployment of the FLUX.1-Kontext-dev model.

💡 Tip: If you want to use cloud API models for image editing, please see Image Editing Pipeline (API Version)

---

2. Quick Start

Step 1: Create a New DataFlow Working Directory

mkdir run_dataflow_mm
cd run_dataflow_mm

Step 2: Configure Model Path

Configure the model path in the pipeline code. Two methods are supported:

(1) Method 1: Use Hugging Face model path (auto-download)

hf_model_name_or_path="black-forest-labs/FLUX.1-Kontext-dev"

(2) Method 2: Use local model path (downloaded model)

hf_model_name_or_path="/path/to/your/local/FLUX.1-Kontext-dev"

Modify the hf_model_name_or_path parameter of LocalImageGenServing in image_editing_pipeline.py:

self.serving = LocalImageGenServing(
    image_io=ImageIO(save_path=image_save_path),
    hf_model_name_or_path="black-forest-labs/FLUX.1-Kontext-dev",  # Model path
    hf_cache_dir="./cache_local",  # Hugging Face model cache directory
    hf_local_dir="./ckpt/models/",  # Local model storage directory
    Image_gen_task="imageedit",  # Image editing task type
    batch_size=4,  # Batch size
    diffuser_model_name="FLUX-Kontext",  # Diffusion model name
    diffuser_num_inference_steps=28,  # Diffusion model inference steps, adjustable to balance speed and quality
    diffuser_guidance_scale=3.5,  # Guidance scale, controls the match between generated results and prompts
)

Step 3: Prepare Image and Text Data

We use jsonl files to store image and text data, with one sample per line. Here is a simple example of input data:

{"images": "image.png", "conversations": [{"role": "user", "content": "Change the vase to red."}]}
{"images": "human_inpaint.jpg", "conversations": [{"role": "user", "content": "Complete the shadowed part into a generic person's portrait."}]}

images is the path to the image to be edited; conversations contains a list of dialogues with image editing instructions, and the content field is the text prompt.

Step 4: Run the Pipeline

python dataflow/statics/gpu_pipelines/image_editing_pipeline.py

Generated files will be saved by default in the ./cache_local/image_edit_local directory.

---

3. Data Flow and Pipeline Logic

1. Input Data

The input data for this pipeline includes the following fields:

• images: Path to the image(s) to be edited, supporting image formats such as png, jpg, etc.
• conversations: Dialogue format data containing image editing instructions.

This input data is stored in jsonl files and managed and read through the FileStorage object:

self.storage = FileStorage(
    first_entry_file_name="<your_jsonl_file_path>",
    cache_path="./cache_local/image_edit_local",
    file_name_prefix="dataflow_cache_step",
    cache_type="jsonl"
)

2. Image Editing Generation (PromptedImageEditGenerator)

The core step of the pipeline is using the Prompted Image Edit Generator (PromptedImageEditGenerator) combined with local GPU models to generate edited images based on input images and editing instructions.

Features:

• Generate edited images from images and editing instructions using local GPU models (e.g., FLUX.1-Kontext-dev)
• Support configuration of inference steps, guidance scale, and other parameters
• Adjustable batch size to optimize GPU utilization
• Automatically save generated images to specified paths

Input: Image file paths and dialogue format data (containing editing instructions)
Output: Edited image file paths

Local GPU Service Configuration:

self.serving = LocalImageGenServing(
    image_io=ImageIO(save_path=image_save_path),  # Image save path
    hf_model_name_or_path="black-forest-labs/FLUX.1-Kontext-dev",  # Model path
    hf_cache_dir="./cache_local",  # Hugging Face model cache directory
    hf_local_dir="./ckpt/models/",  # Local model storage directory
    Image_gen_task="imageedit",  # Task type: image editing
    batch_size=4,  # Batch size
    diffuser_model_name="FLUX-Kontext",  # Diffusion model name
    diffuser_num_inference_steps=28,  # Diffusion model inference steps
    diffuser_guidance_scale=3.5,  # Guidance scale
)

Operator Initialization:

self.image_edit_generator = PromptedImageEditGenerator(
    image_edit_serving=self.serving,  # Image editing service
    save_interval=10  # Save interval
)

Operator Execution:

self.image_edit_generator.run(
    storage=self.storage.step(),
    input_image_key="images",  # Input image field
    input_conversation_key="conversations",  # Input dialogue field
    output_image_key="output_image",  # Output image field
)

3. Output Data

Finally, the output data generated by the pipeline will include the following:

• images: Original input image path(s)
• conversations: Original dialogue data (containing editing instructions)
• output_image: List of edited image file paths

Output Data Example:

{"images":"image.png","conversations":[{"role":"user","content":"Make the whole picture more vibrant."}],"output_image":["./cache_local/image_edit_local/sample_0/sample_0_0.png"]}

---

4. Pipeline Example

Below is an example of an image editing pipeline using local FLUX models:

import os
from pathlib import Path
from dataflow.operators.core_vision import PromptedImageEditGenerator
from dataflow.serving.local_image_gen_serving import LocalImageGenServing
from dataflow.utils.storage import FileStorage
from dataflow.io import ImageIO


class ImageEditingPipeline():
    def __init__(self):
        current_file = Path(__file__).resolve()
        project_root = current_file.parent.parent.parent.parent.parent
        
        data_file = project_root / "dataflow" / "example" / "image_gen" / "image_edit" / "prompts_local.jsonl"
        
        # -------- Storage Configuration --------
        self.storage = FileStorage(
            first_entry_file_name=str(data_file),
            cache_path="./cache_local/image_edit_local",
            file_name_prefix="dataflow_cache_step",
            cache_type="jsonl"
        )

        image_save_path = str(project_root / "cache_local" / "image_edit_local")
        
        # -------- Local GPU Image Editing Service --------
        self.serving = LocalImageGenServing(
            image_io=ImageIO(save_path=image_save_path),
            hf_model_name_or_path="black-forest-labs/FLUX.1-Kontext-dev",  # Or local model path
            hf_cache_dir="./cache_local",
            hf_local_dir="./ckpt/models/",
            Image_gen_task="imageedit",
            batch_size=4,
            diffuser_model_name="FLUX-Kontext",
            diffuser_num_inference_steps=28,
            diffuser_guidance_scale=3.5,
        )

        # -------- Image Editing Generation Operator --------
        self.image_edit_generator = PromptedImageEditGenerator(
            image_edit_serving=self.serving,
            save_interval=10
        )
    
    def forward(self):
        # Call PromptedImageEditGenerator to generate edited images
        self.image_edit_generator.run(
            storage=self.storage.step(),
            input_image_key="images",
            input_conversation_key="conversations",
            output_image_key="output_image",
        )

if __name__ == "__main__":
    # -------- Pipeline Entry Point --------
    model = ImageEditingPipeline()
    model.forward()