Kohya Folder Names & Civitai Zip Layout for LoRA Datasets

You already have paired images and captions—or a LoRA Dataset export—and still stall at the last mile: Kohya GUI wants a weird folder name, Civitai rejects your zip, or the trainer logs “No data found.” This guide owns that packaging contract. It covers folder naming, where to point train_data_dir, caption sidecars, Kohya’s TOML mode, and the flat zip Civitai’s on-site trainer expects. If you are still building the pairs from one reference, start with creating a LoRA dataset from one image; this post picks up at export.

What this guide covers (and what it does not)

Scope here is filesystem layout, caption sidecars, zip root, and the two Kohya loading modes (subdirectory naming vs dataset TOML). Out of scope: learning rate, network dim, epoch choice, caption wording, and image count. You will leave with:

• A correct Kohya DreamBooth-style tree and the parent vs image-folder rule.
• A Civitai zip that keeps captions paired at the archive root.
• A short checklist to go from one export to either trainer without rewriting captions.

For caption content and trigger placement, use the captioning guide. For how repeats interact with image count and steps, see how many images to train a LoRA.

The two packaging contracts (Kohya vs Civitai)

Same captioned pairs can feed both trainers. You rearrange folders or zip contents; you do not re-caption. Caption style (natural language vs tags) still follows the base model—see Flux vs SDXL LoRA training—but the packaging steps below are shared.

Kohya DreamBooth-style folder naming

Anatomy of 10_ohwx woman

In classic DreamBooth-style loading, Kohya scans child folders of train_data_dir and parses each name on the first underscore. The leading integer becomes num_repeats; everything after becomes class_tokens (spaces allowed). Examples from the GUI docs include 30_cat and 40_black mamba. Folders whose names do not start with an integer repeat are ignored (sd-scripts logs a warning).

• 10_ohwx woman → 10 repeats, class tokens ohwx woman
• 30_cat → 30 repeats, class tokens cat

Parent folder vs image folder (the #1 Kohya mistake)

Point train_data_dir at the parent of the N_… folders, not at the folder that literally contains the PNGs. Kohya looks for named children under that path. Aiming one level too deep is the usual cause of No data found. Please verify arguments.

train_data/                    ← point train_data_dir HERE
└── 10_ohwx woman/             ← repeats_trigger class
    ├── 001.png
    ├── 001.txt
    ├── 002.png
    └── 002.txt

Wrong: setting train_data_dir to train_data/10_ohwx woman. Also wrong: dumping images loose in train_data/ with no N_name child—those files are not discovered in DreamBooth-style mode.

Multi-concept layouts follow the same parent rule. From the kohya_ss folder docs you can have images/30_cat and images/30_dog, plus optional regularization siblings such as regularization/1_cat. This post’s hero case is one character concept.

Caption sidecars (.txt / .caption)

Put a caption file beside each image with the same basename: 001.png pairs with 001.txt. When a .txt or .caption exists, its contents take precedence over the folder’s class tokens for that image. Folder class_tokens are only the fallback when no caption file exists. Put the trigger in the caption (see the captioning guide); do not assume the folder name alone teaches the trigger when sidecars are present.

Kohya TOML / --dataset_config mode (folder names ignored)

With an explicit dataset config, image_dir must contain images directly. Naming a folder 5_cat does not apply repeats or class tokens in this mode—set num_repeats and class_tokens explicitly. class_tokens are used only when a caption file is missing; missing both caption and class tokens errors out.

[general]
caption_extension = ".txt"
keep_tokens = 1

[[datasets]]
resolution = 1024
batch_size = 1

  [[datasets.subsets]]
  image_dir = "/path/to/ohwx_woman_images"
  class_tokens = "ohwx woman"
  num_repeats = 10

Packing for Civitai’s on-site LoRA trainer

Filename pairing

Civitai accepts a .zip of images only, or images plus matching caption .txt files. Pairing is same basename (1.jpg + 1.txt). Caption extension is .txt only. Guidelines list .png, .jpg, and .jpeg; the education page also lists .webp. Prefer lowercase extensions. Community reports of uppercase extensions or sequential mis-numbering (01.jpg with 02.txt) breaking caption detection are common enough to avoid.

Zip the files, not the parent folder

Select the image/caption pairs; do not include the parent folder when preparing the zip. Desired layout inside the archive:

dataset.zip
├── 001.png
├── 001.txt
├── 002.png
└── 002.txt

Bad layout (common failure):

dataset.zip
└── my_lora_dataset/
    ├── 001.png
    └── 001.txt

Also unnecessary for the on-site trainer: wrapping pairs in 10_ohwx woman/. Repeats are configured in the Civitai trainer UI, not via Kohya folder names.

Optional: images without captions

Civitai accepts image-only zips and can auto-label (tag vs natural language). Prefer shipping captions from LoRA Dataset so the trigger stays consistent. Match sentence vs tag style to your base model using the captioning guide.

From LoRA Dataset export → Kohya or Civitai

1. 01

   Confirm every image has a same-stem .txt

   Or decide captionless only if you will rely on Civitai auto-label. Matching stems matter more than fancy filenames.
2. 02

   Kohya path: nest under repeats_trigger class

   Create parent train_data/, child {repeats}_{trigger class}/, move pairs inside, point the GUI / train_data_dir at the parent, then verify the steps preview or image count in the logs.
3. 03

   Civitai path: zip only the loose pairs

   Select the files (not the folder), zip, upload, and confirm captions appear under each thumbnail before you submit.
4. 04

   Quick preflight

   No nested junk folders, lowercase extensions, UTF-8 captions, and no leftover Copy of … files.

Doing this with LoRA Dataset

LoRA Dataset exports paired image + .txt captions from one reference (free with your own Gemini API key). After download, nest the pairs under a {repeats}_{trigger class} folder for Kohya, or zip the flat pairs for Civitai. No re-captioning is required if the trigger is already in the sidecars.

Troubleshooting

If packaging looks correct but likeness still wanders at generation time, the problem is probably the dataset or training settings—see fixing identity drift.