CXH_IC_lora_reversal

Here's the trick this pack is quietly built around: IC-LoRA captioning only writes captions in one direction ("here's the garment, here's it being worn"). CXH_IC_lora_reversal manufactures the reverse - captions that describe the same pair going the other way ("here's the model wearing it, here's the garment") - without re-running the vision model at all. You get a second, direction-reversed training set for free, which materially helps a LoRA generalize rather than just memorizing one arrangement.

It's a text-processing node wearing a caption node's clothes. No model input, no image captioning happens - it reads the captions you already generated, flips the framing, and writes new pairs.

How it works

For each image in dir1, it needs:

1. The paired image in dir2 (same filename),
2. An existing caption .txt in text_dir (same base filename).

It reads the existing caption, slices out the segment between slic_start (default [image1]) and slic_end (default [image2]) - i.e. the actual model-written description - and feeds it into a reversed template whose default reads: Realistic style, [cloth-on], the image pair highlights a transformation from a female model wearing the cloth to its clothing sample photo. [image1] a female model is wearing a cloth with {caption} [image2] the clothing sample photo of what the model is wearing in [image1] with {caption}.

The rewritten caption plus a re-stitched image pair (same direction/match_image_size logic as the other IC nodes) land in save_dir. If the output file already exists, it skips - so re-runs are cheap and resume-friendly.

Inputs

• dir1, dir2 - the image pair folders.
• text_dir - where the source .txt captions live.
• save_dir - reversed output goes here.
• slic_start, slic_end - the markers around the caption you want to reuse; default to [image1] and [image2], matching the templates used by the captioning nodes.
• text1, text2, template, direction, match_image_size, format.
• Output: caption (STRING) - last processed caption; real output is on disk.

Installing

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

No model downloads - this node never touches a model.

Gotchas

• Missing pair or missing .txt → skipped silently. It prints "Second image not found" but there's no error dialog. Check your file counts.
• The slice assumes your source captions contain [image1] and [image2] markers. If your captioner ran with a template that omitted them, .index() raises an error. Use the pack's own IC caption nodes upstream so the markers exist.
• Skipping existing files is the resume feature - but if you changed the template, stale files won't be regenerated. Clear save_dir when you change anything.