📚 LoRA Dataset Forge

📚 FiL LoRA Dataset Forge — turns an image batch into a training-ready dataset: aspect-ratio buckets at the chosen resolution, one LLM caption per image, and a kohya_ss/sd-scripts (or flat) folder written to disk with a manifest.

By FiL-Design-Ai·Created 2 months ago·Updated 16 days ago· 3

image
config

preview
report
dataset_path
manifest

◄dataset_namemy_lora►

◄trigger_word►

◄class_token►

◄base_resolution1024►

◄layoutkohya►

◄repeats10►

◄caption_modenatural►

◄crop_modecenter►

◄dry_runfalse►

◄write_modeappend►

◄caption_max_words60►

◄caption_languageen►

◄bucket_step64►

◄caption_extension.txt►

◄image_formatpng►

◄jpg_quality95►

◄seed-1►

◄captions►

◄dont_caption►

◄caption_instruction►

Category🎨 FiL Design/Dataset

Inputs (22)

Name	Type	Default	Description
image	IMAGE		Images to turn into a dataset. The whole batch is processed, one file per frame.
dataset_name	STRING	my_lora	Dataset folder name under ComfyUI/output/datasets. Sanitized to a safe path segment.
trigger_word	STRING		Unique token that activates the LoRA, prepended to every caption (e.g. 'ohwx'). Leave empty for a style LoRA.
class_token	STRING		Class the subject belongs to (e.g. 'woman', 'car'). Follows the trigger word in captions and in the kohya folder name.
base_resolution	COMBO	1024	Training resolution. Buckets are built around this area — 1024 for SDXL/Flux, 512-768 for SD 1.5.
layout	COMBO	kohya	kohya: <name>/img/<repeats>_<trigger> <class>/ plus dataset.toml. flat: images and captions directly in <name>/.
repeats	INT	101–1000	Repeats per image per epoch — encoded in the kohya folder name and dataset.toml. Fewer images means more repeats.
caption_mode	COMBO	natural	natural: prose for Flux/SDXL. tags: comma tags for SD 1.5/Pony. hybrid: short comma phrases. none: trigger word only, no LLM call.
crop_mode	COMBO	center	How the frame is cropped into its bucket. entropy keeps the most detailed region instead of the middle.
dry_run	BOOLEAN	false	Compute buckets, captions and file names but write nothing. Read the report to check the plan first.
write_mode	COMBO	append	append: keep existing files and continue numbering. overwrite: delete this node's image/caption pairs in the folder first (other files are left alone).
caption_max_words	INT	604–400	Hard word limit per caption. Captions longer than this are truncated.
caption_language	COMBO	en	Caption language. Trainers expect English unless the base model was trained otherwise.
bucket_step	INT	648–256	Bucket side divisor. 64 is the kohya default and fits SD/SDXL/Flux latents.
caption_extension	COMBO	.txt	Extension of the caption sidecar files. kohya defaults to .txt.
image_format	COMBO	png	png keeps the crop lossless; jpg saves disk space on large sets.
jpg_quality	INT	9550–100	JPEG quality when image_format is jpg. Below ~90 the artifacts start training too.
seed	INT	-1-1–999999999999	Provider-side caption seed. -1 lets the provider pick one; a fixed value makes captions reproducible.
configopt	FIL_PROVIDER_CONFIG		Provider Loader config. Required for LLM captioning — leave unconnected when captions come from the captions input or caption_mode is 'none'.
captionsopt	STRING		Manual captions, one per image, separated by a line containing only '---'. Accepts Optic Scanner output as-is. When filled, no LLM call is made.
dont_captionopt	STRING		Invariants the captions must never mention (e.g. 'her face, red hair'). These belong to the trigger word — describing them stops the LoRA from learning them.
caption_instructionopt	STRING		Extra instruction appended to every caption request.

Outputs (4)

Name	Type	Description
preview	IMAGE	Bucketed frames letterboxed onto a square canvas so the batch can be previewed. The files on disk keep their real bucket size.
report	STRING	Human-readable summary: buckets, captions, warnings, target path.
dataset_path	STRING	Absolute path of the dataset folder that was written.
manifest	DICT	Full manifest dict — per-image bucket, crop, caption and hash.