joy diretory caption

Name aside ("diretory" - a typo the author kept), this is the workhorse of the ymc_node_joy pack. Point it at a folder of images and it captions every one of them with Joy Caption, one pass through the model, and optionally drops a .txt file next to each image. That's the entire LoRA-dataset pipeline: images in, natural-language captions out, in exactly the format kohya, ai-toolkit and OneTrainer expect.

Why you'd reach for it

If you're training a LoRA on an LLM-encoder model - Flux, Qwen-Image, Z-Image - your training captions should be natural language, not tags. Joy Caption is the community's go-to for writing those, and the KB's training guidance is blunt: captioning quality beats dataset size, and BLIP is not an acceptable substitute. This node is how you caption 50 images without clicking through a single one. Set it up, run it, go make tea - each image takes a couple of seconds plus the model load, depending on your GPU.

How it works

You feed it the JoyPipeline from the joy model loader and an image_dir path. It lists the images in that folder (only .png, .jpg, .jpeg, .webp, and only the top level - no recursion), loops through them with the model already loaded, and produces a caption per image. If an image fails to open, it logs the error, writes a ,,, placeholder, and keeps going - one bad file won't kill your batch.

Two outputs come back: image_files (the full paths, one per line) and captions (the matching captions, one per line). If you turned on save_caption, you'll also find imagename.txt files sitting right next to the source images.

The inputs that matter

• joy_pipeline - from the model loader. No loader, no captions.
• image_dir - the absolute path to your folder. This is the one beginners mangle; see below.
• trigger - your LoRA trigger word, prepended to every caption as trigger, <caption>. This is huge for training: a trigger word makes the captioner include your concept token so it doesn't get baked into the model as background.
• save_caption - write the .txt files. You want this on for real training runs.
• prompt - the instruction the LLM follows, like "A descriptive caption for this image". Defaults to a plural version since you're captioning a batch.

The same cache toggle exists here, but it's far less of a footgun: the directory node clears the pipeline only after the whole folder is done, so a batch runs in one model load regardless. Leave it off.

Where people get burned

The most common failure is Directory not found - the node takes a plain filesystem path, not a ComfyUI output directory, so paste the full path (/home/you/images/training_set, or C:\Users\... on Windows). Watch for the folder-not-recursive behavior: subfolders are silently ignored. And after a run, scan the generated captions for ,,, - those are the images that errored, and you want to know before you train on a dataset with holes in it.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/YMC-GitHub/ymc_node_joy

Restart ComfyUI and let the deps auto-install; or grab it via ComfyUI Manager under "ymc_node_joy". Model setup is shared with the rest of the pack - the loader auto-downloads SigLIP and your chosen Llama on first run, but you must drop models/Joy_caption/image_adapter.pt in place yourself or nothing loads.