BILL-Concept_Isolator-Captioner

Overview

The BILL-Concept_Isolator-Captioner is a specialized ComfyUI node designed to automate the creation of high-precision captioning datasets. By integrating multimodal LLMs—specifically the Gemma 4 architecture—it bridges the gap between raw image data and the descriptive metadata required for high-quality LoRA (Low-Rank Adaptation) training.

Technical Approach: Concept Isolation

To ensure maximum subject fidelity, the node employs a logic gate known as Concept Isolation. This process prevents the model from overfitting by decoupling the subject's unique identifiers from the general scene description.

The Workflow Logic:

• Input Filtering: The user specifies the subject's inherent traits (Excluded Details).
• Constraint Application: The LLM is prompted to describe the image while treating the excluded details as "invisible" to the general description.
• Token Injection: The identified unique traits are condensed into a single Trigger Word, ensuring the LoRA associates the trigger token with the visual identity.

System Capabilities

• High-Fidelity Descriptive Analysis: Focuses on lighting, composition, and texture rather than simple object labeling.
• Single-Pass Inference: Reduces token drift by utilizing a single, complex prompt pass rather than iterative refinement.
• Integrated File Management: Streamlined .txt export mapped to image filenames for immediate use in training scripts (e.g., AI Toolkit).
• Hardware Target: Optimized for high-VRAM environments (24GB+), enabling the use of larger, more reasoned models.

Installation Guide

Backend Setup

1. Install the Ollama runtime environment.
2. Deploy the vision model via CLI: ollama run gemma4

Node Integration

1. Clone the repository to the ComfyUI custom nodes directory: git clone https://github.com/0-bill-0/ComfyUI-BILL-Concept_Isolator-Captioner.git
2. Install required Python dependencies: pip install -r requirements.txt

Hardware Requirements

| Component | Minimum | Recommended | | :--- | :--- | :--- | | VRAM | potato | 24GB - 32GB | | Model | potato | Gemma 4 | | OS | potato | Linux |