Sign In

Kiko dataset tagger (Qwen-VL captions for LoRA training)

Download

1 variant available

Archive Other

kiko-dataset-tagger-v6.1.zip

10.43 KB

Verified:

Type
Workflows
Stats

37

Reviews
Published

Nov 9, 2026

Base Model

Other

Hash
AutoV209C0856FF9
kiko-dataset-tagger-workflow.png

Captions every image in a dataset folder for LoRA training and writes <image name>.txt next to each one. Pick the folder by browsing in the Kiko Local Image Loader gallery, set your trigger word, pick the base model you will train in the caption prompt studio, and queue. The caption prompts come from Mick Mumpitz's CCC Prompt Studio, and the vision model is Huihui-Qwen3-VL-8B-Instruct-abliterated (QwenVL). Captions are only written inside a folder named datasets that WAS Node Suite is allowed to write to. Add it once to paths.allow_write in ComfyUI/user/was-node-suite/config.yaml and restart ComfyUI: ["/home/you/training/datasets"] on Linux/macOS, ["D:/training/datasets"] on Windows (forward slashes). Works on Windows, Linux and macOS.

Needs ComfyUI-QwenVL 2.3.2 or newer. Its downloader node gained a setting in 2.3.2; on older releases the Download models node's values land in the wrong boxes. Update QwenVL in ComfyUI Manager, then reload the workflow.

v6.1: caption file names work with Windows paths and any parent folder (v6 only matched the folder in my file), the old QwenVL version stamp is gone, and the note has the Windows setup.

Files: kiko-dataset-tagger-v6.1.json

Custom Nodes

Models

  • Huihui-Qwen3-VL-8B-Instruct-abliterated (folder, about 17 GB) → models/LLM/huihui-ai/: https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated (the Download models group fetches it with QwenVL's downloader and registers it in QwenVL's custom_models.json)

  • Optional alternatives in the list: Huihui-Qwen3.5-9B-abliterated, Huihui-Qwen3.5-4B-abliterated (huggingface.co/huihui-ai)

First run

Each workflow has a bypassed Download models group. Turn it on in Toggle stages, queue once, then turn it off again. Files you already have are skipped. Civitai models and gated Hugging Face models need a token, set where ComfyUI starts (never in the workflow):

# Linux / macOS (~/.bashrc, ~/.zshrc or your start script)
export CIVITAI_API_KEY="your-civitai-key"
export HF_TOKEN="hf_your-token"

# Windows (Command Prompt once, then a new window)
setx CIVITAI_API_KEY "your-civitai-key"
setx HF_TOKEN "hf_your-token"

Restart ComfyUI after setting them.

Workflow image

No example render: this workflow writes text, not images.

kiko-dataset-tagger-workflow.png is a picture of the whole graph (v6; the v6.1 changes are in the regex and note text only). Load the .json to use it.

Happy rendering! Kiko