ComfyUI-Omini-Kontext

Wrapper ComfyUI integration for the Flux Omini Kontext pipeline, enabling seamless character/object insertion into scenes using FLUX.1-Kontext-dev with LoRA adaptation.

Features

• Character/Object Insertion: Insert reference images into scenes with precise spatial control
• LoRA Support: Load and use pre-trained LoRA weights for specific insertion tasks
• Memory Optimization: Built-in VAE slicing and tiling for efficient VRAM usage
• Flexible Pipeline: Support for both text-to-image and image-to-image workflows
• Position Control: Fine-tune object placement with reference_delta parameters

Installation

1. Clone the repository into your ComfyUI custom_nodes folder:

   cd ComfyUI/custom_nodes
   git clone https://github.com/tercumantanumut/ComfyUI-Omini-Kontext.git
   

2. Install dependencies:

   cd ComfyUI-Omini-Kontext
   pip install -r requirements.txt
   

3. Download the base model (optional - will auto-download on first use):

   • The pipeline uses black-forest-labs/FLUX.1-Kontext-dev by default
   • Requires HuggingFace login: huggingface-cli login

4. Download pre-trained LoRA weights (optional):

   # Example: Character insertion LoRA
   wget https://huggingface.co/saquiboye/omini-kontext-character/resolve/main/character_5000.safetensors \
        -O ComfyUI/models/loras/omini_kontext_character_5000.safetensors
   

Available Nodes

1. Omini Kontext Pipeline Loader

Loads the Flux Omini Kontext pipeline with optional LoRA weights.

• Inputs:
  • model_path: HuggingFace model ID or local path
  • lora_path: Optional path to LoRA weights
• Output: OMINI_KONTEXT_PIPELINE

2. Omini Kontext Pipeline

Main generation node for character/object insertion.

• Required Inputs:
  • pipeline: Loaded pipeline from loader node
  • prompt: Text description
  • reference_image: Character/object to insert
  • reference_delta_x/y/z: Position control (default: 0, 0, 96)
  • Generation parameters (steps, guidance_scale, width, height, seed)
• Optional Inputs:
  • input_image: Base image for img2img mode
  • negative_prompt: Negative text prompt
  • true_cfg_scale: Additional CFG control
• Output: Generated image

3. Omini Kontext Image Scale

Scales images to optimal Kontext resolutions.

• Input: Any image
• Output: Scaled image at optimal resolution

4. Omini Kontext LoRA Loader

Load LoRA weights into an existing pipeline.

• Inputs:
  • pipeline: Pipeline to add LoRA to
  • lora_name: LoRA file from models/loras folder
  • strength: LoRA strength multiplier
  • adapter_name: Name for the adapter

5. Advanced Encoder Nodes

For advanced workflows:

• Image Encoder: Encode images to latents
• Text Encoder: Encode prompts to embeddings
• Reference Encoder: Encode reference with position delta
• Latent Combiner: Combine input and reference latents

Basic Workflow

1. Load Pipeline: Use "Omini Kontext Pipeline Loader" with model path
2. Load LoRA (optional): Use "Omini Kontext LoRA Loader"
3. Prepare Images:
   • Load your input image (optional)
   • Load your reference character/object image
   • Optionally scale with "Omini Kontext Image Scale"
4. Generate: Connect everything to "Omini Kontext Pipeline" node
5. Save Result: Use standard ComfyUI save image node

Example Use Cases

Character Insertion

Insert a specific character into various scenes:

reference_delta = [0, 0, 96]  # Standard positioning
prompt = "A boy playing in a sunny park"

Object Placement (If trained.)

Place objects with spatial control:

reference_delta = [50, 0, 96]  # Shift right
prompt = "A vintage car parked on a city street"

Style Transfer (If trained.)

Combine reference style with scene:

reference_delta = [0, 0, 48]  # Closer integration
prompt = "In the style of the reference"

Tips

1. Reference Delta Values:

   • X: Horizontal position (-100 to 100 typical)
   • Y: Vertical position (-100 to 100 typical)
   • Z: Depth/integration (48-144 typical, 96 default)

2. Memory Management:

   • Pipeline automatically enables VAE slicing/tiling
   • For 24GB VRAM: up to 1024x1024 generation
   • For 16GB VRAM: recommended 768x768 or lower

3. LoRA Strength:

   • 1.0 = full strength (default)
   • 0.5-0.8 = subtle effect
   • 1.2-1.5 = stronger effect

Troubleshooting

"No module named 'diffusers'"

Run: pip install git+https://github.com/huggingface/diffusers.git

"CUDA out of memory"

• Reduce generation resolution
• Close other GPU applications
• Enable CPU offloading (future feature)

"401 Unauthorized" when loading model

Run: huggingface-cli login and enter your HuggingFace token

Credits

• Original Omini-Kontext implementation: Saquib764/omini-kontext
• Based on FLUX.1-Kontext-dev by Black Forest Labs
• ComfyUI integration by ogkai (github: tercumantanumut)

License

This project follows the same license as the original omini-kontext repository.