This workflow automatically generates clean, structured image captions by combining WD14 tagging, Florence-style natural language descriptions, and a custom trigger token for training consistency. The idea behind this workflow is to deliver proven results for easily (one-click) captioning datasets for training. I have made many high quality LORA from the datasets this workflow outputs.

Designed for:

• Dataset preparation

• LoRA / model training

• Caption inspection & QA

• Prompt reconstruction workflows

The examples shown are direct outputs from the workflow, displayed exactly as generated. No LLM or API keys needed. If you’re tired of messy captions or inconsistent datasets, this keeps everything clean, readable, and repeatable in one easy simple workflow.

What This Workflow Does

• Uses WD14 to extract high-quality tag metadata

• Uses Florence to generate a natural-language image description

• Injects a custom trigger token at the start of every caption

• Outputs both tags + descriptive text in a single caption block

• Saves captions to a user-defined folder inside ComfyUI/output

This keeps captions:

• Consistent

• Human-readable

• Training-friendly

📁 Important Setup Note (VERY IMPORTANT)

You must create a folder inside:

ComfyUI/input/

Example:

ComfyUI/input/Captions

Then select that folder in the caption loader node. If the folder does not exist, the workflow will not write captions correctly as it doesn't know which directory to load from.

About the Example Images

Each example image shows:

• The original image

• A caption panel

Nothing is manually edited or rewritten — this is raw workflow output for transparency and accuracy.

Example Output Structure

Captions follow this format:

TRIGGER, wd14_tags_here,
florence_generated_description_here

This structure is ideal for:

• LoRA trigger anchoring

• Dataset reuse

• Easy downstream parsing

Recommended Use Cases

• LoRA training datasets

• SDXL / Flux / Wan caption generation

• Dataset cleanup & validation

• Prompt reverse-engineering

• Caption benchmarking

Requirements

• ComfyUI (recent build)

  • Must include comfy-core ≥ 0.3.75

  • Required for BETA Dataset nodes:

    • LoadImageDataSetFromFolder

    • SaveImageTextDataSetToFolder

    • StringConcatenate

Custom Node Packs

• comfyui-florence2 (v1.0.6)

• comfyui-wd14-tagger (v1.0.0)

• comfyui-custom-scripts (pysssss) (v1.2.5)

• ComfyUI-Jjk-Nodes

Models

• Florence-2-SD3-Captioner

  • Repo: gokaygokay/Florence-2-SD3-Captioner

• WD14 Tagger Model

  • wd-convnext-tagger-v3

• Folder Setup

  • Create a folder in ComfyUI/input/

  • Example used in workflow: ComfyUI/input/Captions