comfyui-lora-dataset-tools

Streamlined LoRA dataset generation for ComfyUI — two nodes that replace 5+ manual configuration steps.

Before this package you had to edit a base description node, a trigger word in the wildcard node, two separate output folder fields, maintain external .txt wildcard files, and then run a post-processing Python script after every generation. Now you edit one node (LoRADatasetConfig) and everything else wires up automatically. The caption sidecar .txt file is written atomically alongside each PNG — no post-processing script needed.

---

Quick Start

1. Install the package (see Installation).
2. Open an example workflow from example_workflows/.
3. Edit only the 🎛️ LoRA Dataset Config node: set your trigger_word, output_name, and description.
4. Set seed to Randomize for variety across runs.
5. Queue the prompt. Paired .png + .txt files land in ComfyUI/output/lora_dataset/<name>/.

---

Nodes

🎛️ LoRA Dataset Config (LoRADatasetConfig)

The single configuration node. You only ever edit this node. All other nodes in the workflow are pre-wired and never need touching.

Built-in wildcard lists cover shot types, poses, expressions, locations, lighting, and atmosphere — no external .txt files required. The variety is seeded so results are reproducible when you need them and random when you don't.

Inputs

| Name | Type | Default | Description | |------|------|---------|-------------| | trigger_word | STRING | char_trigger | Unique LoRA trigger word. Used in training captions and prepended to the varied prompt. | | output_name | STRING | my_character | Output folder and filename prefix. Use snake_case; spaces are converted automatically. | | description | STRING | adult woman, cel-shaded… | Base description of your subject. Do not include the trigger word here — it is added automatically. | | dataset_type | ENUM | character | character · outfit · location · object. Controls wildcard selection and output resolution. | | seed | INT | 0 | Set to Randomize in the widget for a different prompt composition every run. |

Outputs

| Name | Type | Description | |------|------|-------------| | trigger | STRING | Trigger word, trimmed. Connect to LoRACaptionSaver.trigger_word. | | varied_prompt | STRING | trigger, <random shot/pose/expression/location/lighting>. Wire to your positive prompt. | | base_description | STRING | The description field verbatim. Use as a stable secondary prompt or for IPAdapter conditioning. | | output_prefix | STRING | lora_dataset/<name>/<name>. Connect to LoRACaptionSaver.filename_prefix. | | latent_width | INT | 1216 for location, 1024 for all others. Wire to your Empty Latent Image node. | | latent_height | INT | 832 for location, 1024 for all others. Wire to your Empty Latent Image node. |

Resolution by dataset type

| dataset_type | Width | Height | Notes | |----------------|-------|--------|-------| | character | 1024 | 1024 | Full IPAdapter face chain in example workflow | | outfit | 1024 | 1024 | Single IPAdapter PLUS in example workflow | | location | 1216 | 832 | Landscape ratio, no people | | object | 1024 | 1024 | Isolated subject |

Built-in wildcard categories

The wildcards are baked into wildcards.py — no .txt files to maintain.

| Category | Used for | |----------|----------| | char_shot | Camera framing (full body, bust, close-up, etc.) | | char_pose | Body language and stance | | char_expression | Facial expression | | location | Background environment | | lighting | Lighting style | | outfit_shot | Outfit-specific framings | | outfit_activity | Activity showing outfit in motion | | obj_angle | Viewing angle for locations/objects | | time_of_day | Daylight condition | | atmosphere | Mood and weather |

---

💾 LoRA Caption Saver (LoRACaptionSaver)

Saves each generated image as a PNG and writes a matching .txt caption file in the same atomic operation. Replaces the combination of SaveImage + SaveText|pysssss + a separate post-processing script.

Inputs

| Name | Type | Required | Description | |------|------|----------|-------------| | images | IMAGE | Yes | Image tensor from your KSampler/VAE Decode pipeline. | | caption | STRING | Yes | Caption string from Florence2Run (or any STRING node). Force-input — must be connected. | | filename_prefix | STRING | Yes | Output path prefix. Connect to LoRADatasetConfig.output_prefix for automatic naming, or type lora_dataset/<name>/<name> manually. | | trigger_word | STRING | No | Connect to LoRADatasetConfig.trigger. Prepended to the caption in the .txt file. |

Outputs

| Name | Type | Description | |------|------|-------------| | image | IMAGE | Pass-through of the input image. Chain to a preview node if desired. | | caption | STRING | Final caption string as written to the .txt file (trigger_word, <Florence2 caption>). |

Caption file format

my_character, The image shows a woman standing confidently in a forest clearing...

If no trigger_word is connected the caption is written as-is without a prefix.

---

Workflow Guide

Step-by-step

1. Open an example workflow matching your subject type (character / outfit / location-object).
2. Edit 🎛️ LoRA Dataset Config — the only node you need to change:
   • trigger_word: your unique activation token (e.g. jane_doe_v1)
   • output_name: short snake_case name (e.g. jane_doe)
   • description: visual description of your subject without the trigger word
   • dataset_type: pick the type that matches your subject
   • seed: set to Randomize for maximum variety across the dataset
3. Load reference images into the LoadImage node(s) (face reference for character/outfit, style reference for others).
4. Queue the prompt (Ctrl+Enter). ComfyUI will generate the image, auto-caption it with Florence2, and save the pair.
5. Repeat — queue 20–50 times with seed on Randomize. All pairs accumulate in ComfyUI/output/lora_dataset/<name>/.

The dataset folder is ready for training immediately after generation — no post-processing, no script to run.

---

Example Workflows

All example workflows are in the example_workflows/ directory.

| File | Type | Description | |------|------|-------------| | 21_char_lora_dataset.json | Character | 3-stage IPAdapter face identity chain for strong character consistency | | 22_outfit_lora_dataset.json | Outfit | Single IPAdapter PLUS for style and outfit consistency | | 23_loc_obj_lora_dataset.json | Location / Object | Landscape ratio (1216×832), no IPAdapter, environment focus |

To use: in ComfyUI open the workflow browser, drag-and-drop the .json file, or use Load from the menu.

---

Output Format

Generated files are saved under ComfyUI/output/:

ComfyUI/output/lora_dataset/
└── my_character/
    ├── my_character_00001_.png
    ├── my_character_00001_.txt
    ├── my_character_00002_.png
    ├── my_character_00002_.txt
    ├── my_character_00003_.png
    ├── my_character_00003_.txt
    └── ...

Each .txt file contains a single line:

my_character, The image shows a woman in a full body shot, standing confidently with hands on hips, in a forest clearing, soft natural daylight

The PNG files embed the full ComfyUI workflow metadata (prompt, nodes) in their EXIF data.

---

Training Preparation

The output folder is structured exactly as kohya_ss and SimpleTuner expect — each image paired with a same-name .txt caption file.

kohya_ss

Point your dataset directory to ComfyUI/output/lora_dataset/<name>/. The .txt sidecar files are picked up automatically as captions.

SimpleTuner

Use the same directory as your instance_data_dir. Set caption_strategy: textfile in your dataset config — SimpleTuner will read the .txt files alongside each image.

Recommended dataset sizes

| Type | Images | |------|--------| | Character | 20–50 | | Outfit | 15–30 | | Location | 15–25 | | Object | 15–30 |

---

Requirements

Custom nodes (required for example workflows)

| Package | Source | Used for | |---------|--------|----------| | ComfyUI IPAdapter Plus | cubiq/ComfyUI_IPAdapter_plus | Face/style consistency | | comfyui-florence2 | kilic-ai/comfyui-florence2 | Auto-captioning (Florence2Run) | | ComfyUI-Custom-Scripts | pythongosssss/ComfyUI-Custom-Scripts | ShowText\|pysssss caption preview |

LoRADatasetConfig and LoRACaptionSaver themselves have no extra Python dependencies — all wildcard logic is self-contained in wildcards.py.

Models

| Model | Where to get | |-------|-------------| | ip-adapter-plus_sdxl_vit-h.safetensors | h94/IP-Adapter on HuggingFace → sdxl_models/ | | ip-adapter-plus-face_sdxl_vit-h.safetensors | same repo | | CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors | same repo → models/image_encoder/ | | Florence-2-base | microsoft/Florence-2-base on HuggingFace — place in models/LLM/Florence-2-base/ | | bridgeToonsMix_v80.safetensors | CivitAI — place in models/checkpoints/ |

---

Installation

One-command install (recommended)

bash <(curl -fsSL https://raw.githubusercontent.com/LordTaylor/comfyui-lora-dataset-tools/main/install.sh) /path/to/ComfyUI

Downloads all custom nodes + IPAdapter/CLIP Vision/Florence-2 models automatically. The bridgeToonsMix checkpoint must be downloaded manually from CivitAI.

Manual

cd ComfyUI/custom_nodes

# 1. This package
git clone https://github.com/LordTaylor/comfyui-lora-dataset-tools.git

# 2. Required extensions
git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus.git
git clone https://github.com/kilic-ai/comfyui-florence2.git
git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts.git

Restart ComfyUI. The two nodes appear under the LoRA Dataset Tools category.

ComfyUI Manager

Search for LoRA Dataset Tools in the Manager node browser and click Install. (Available once the package is published to the registry.)

---

License

MIT — see LICENSE.