Joy_caption_alpha_batch_Dirs

If Joy_caption_alpha_batch automates the boring part of one dataset, this node automates the boring part of many. Joy_caption_alpha_batch_Dirs walks every subfolder under img_dir, captions each one with JoyCaption Alpha Two, and - here's the trick - pulls the trigger word out of the folder name itself. The README's rule: name your folders 名字_trigg (name_trigger), e.g. blue_jacket_ohwx. The node splits on the underscore and uses the second part as the activation word for everything in that folder.

This is a quality-of-life node for people managing a whole wardrobe of clothing LoRAs or a batch of character sets at once, where every subfolder is its own training dataset with its own trigger. Instead of running the single-folder batch once per dataset and manually retyping the trigger each time, you name the folders right and hit Run once.

How it works

For each subdirectory of img_dir it runs the same Alpha Two captioning loop as Joy_caption_alpha_batch, but:

• trigger comes from get_trigger_from_string(subdir) - split the folder name on _, take the second part. mydataset_xyz → trigger xyz.
• Output lands in save_dir/<subdir>/, preserving the per-dataset structure.
• Files are written as trigger_<n>.txt (containing trigger,<caption>) and trigger_<n>.<format>.

If the folder name has no underscore, no trigger is written - captions still get saved, just without the activation prefix. Keep your folder names consistent if you want every caption to carry the trigger.

Inputs

• JoyPipeline_alpha - from Joy_caption_alpha_load.
• img_dir - parent folder; each subfolder is treated as a dataset.
• save_dir - mirrored output structure, one subfolder per dataset.
• prompt - default descriptive, or wire in a Joy_caption_alpha_prompt for styled captions.
• format (png/jpg), max_new_tokens, temperature, cache, low_vram, seed - same semantics as the single-folder batch: set cache=True for a real multi-folder run, low_vram=True if the captioner is OOMing next to your checkpoint.

Output is a STRING that's effectively a completion marker; the real work lands on disk.

Installing

Same pack install - ComfyUI Manager (search "Comfyui_CXH_joy_caption") or:

cd ComfyUI/custom_nodes
git clone https://github.com/StartHua/Comfyui_CXH_joy_caption
cd Comfyui_CXH_joy_caption
python -m pip install -r requirements.txt

Restart, and make sure the Alpha Two checkpoint is at models/Joy_caption_alpha/ (manual download - see Joy_caption_alpha_load).

Gotchas

• No underscore, no trigger. The node quietly captions without the activation word, and you may not notice until you're reviewing captions later.
• Empty subfolder raises FileNotFoundError. Check your dataset tree before a long run.
• cache=False default means the model reloads between images - on a many-folder job that's hours of waiting. Flip it on.
• NVIDIA only, as with the whole pack.