Build a LoRA Training Dataset: Sources, Cleanup, Captioning Tools

For related workflows, see LoRA training parameter optimization, dataset bucketing strategies, and LoRA concept design.

TL;DR

Sources, Cleanup, Captioning, Key fixes

• Sources: Danbooru/Safebooru/Gelbooru for anime, Pixiv/DeviantArt/Twitter for artist-specific, Spriters Resource/Creative Uncut for game sprites.
• Cleanup: GIMP for targeted fixes, Experte for AI object removal, Birme for bulk cropping, web tools (WatermarkRemover → Imglarger → Imgupscaler) in incognito mode.
• Captioning: JoyCaption Alpha Two (9 types, 17 options) for descriptive captions; WD14 tagger for tag-style labels; manual review for critical images.
• Key fixes: Clear cookies between free-tool sessions; use JoyCaption's immutable-attribute filter for character LoRAs; run watermark removal first.

Where to Find Training Images

Booru Sites: Primary Source for Anime-Style Data

Booru-style image boards are the most common starting point for LoRA training data. Danbooru, Safebooru, and Gelbooru all use Danbooru's tagging taxonomy, which lets you search by character, artist, style, and conceptual tags simultaneously. Gelbooru is functionally similar to the first two, while SankakuComplex uses a comparable tagging system but carries more mature-themed content.

Artist Platforms and Social Media

Pixiv — the Japanese artist platform — organizes works primarily by series and character rather than by visual features. A Chrome extension called the Powerful Pixiv Downloader streamlines batch collection. DeviantArt hosts a wide range of styles and occasional 3D model packs, though image quality varies considerably. For gathering data from specific artists on Twitter/X, the Twitter Media Downloader extension captures entire artist pages, while Buhitter enables tag-based search beyond what the platform's own interface offers. Zerochan provides additional anime-tagged imagery with a smaller but distinct tag set.

Game Artwork and Sprites

For retro game sprites and console artwork, Spriters Resource organizes material by console and franchise. Creative Uncut provides a secondary source of game-related visual material, useful for fantasy or pixel-art LoRAs.

How to Clean and Prepare Images

Manual Editing

For targeted fixes — removing a single stray line, cropping a white border, or touching up a small artifact — GIMP remains the go-to free editor. A 4-step decensor guide demonstrates how to combine GIMP with Stable Diffusion itself for inpainting-style cleanup.

AI-Powered Object Removal

Experte's Image Cleaner lets you highlight unwanted elements and the AI blends the surrounding context, preserving original resolution. This works well for removing logos, signatures, or small artifacts across large batches.

Bulk Operations

Pinetools handles batch flipping and basic transformations. Birme excels at mass cropping: enter target dimensions and it trims all images to that size automatically.

Free Web Tools for Upscaling and Cleanup

Free web tools for upscaling, denoising, and watermark removal operate per session. Use them in incognito mode and delete cookies between sessions to reset limits. The recommended order is:

Captioning Tools for Dataset Preparation

JoyCaption Alpha Two generates training captions through an open-weight vision-language model. Released in September 2024, it expanded from Alpha One's three caption types to nine: Descriptive, Descriptive (Informal), Training Prompt, MidJourney, Booru tag list, Booru-like tag list, Art Critic, Product Listing, and Social Media Post. The demo offers length selection and 17 extra options that append to the caption prompt.

The developer tested Alpha Two against 15 validation images, scoring +1/-1 detail accuracy against ~200-word human captions. Alpha Two averaged 22.8 vs. GPT-4o's 22.7 with higher variance (SD 5.32 vs. 3.88). The developer describes this as "punching toe-to-toe" with GPT-4o while losing on consistency. The release also fixed a bug where captions near the 256-token limit would truncate mid-sentence.

For LoRA training, a character-name feature lets you specify the name for character references. Extra options let you exclude immutable attributes like ethnicity or gender while including mutable ones like hairstyle.

Important limitation: Training Prompt, Booru tags, and MidJourney modes are unreliable. Informal style captions remain inconsistent.

WD14 Tagger Extension

The WD14 tagger is an extension for Automatic1111's WebUI that applies automated tags to images in bulk. It supports a dropdown with multiple models and can process entire directories.

Note: The GitHub repository previously returned a 404 error. Verify current location before relying on it.

Manual Captioning

Manual captioning remains the most accurate method for small datasets. The process combines visual description with tag-style keywords. For teams, a practical workflow is to use JoyCaption Alpha Two for a first pass, then manually review and correct captions for the most important images.

Common Errors and Fixes

Frequently Asked Questions

What is the best source for anime-style training images? The Danbooru tag network — including Danbooru, Safebooru, Gelbooru, and SankakuComplex — provides the most structured tagging system for concept-based image searches. Using multiple sources increases your usable pool.

How do I caption images automatically for LoRA training? JoyCaption Alpha Two is an open-weight model that generates descriptive captions through a Hugging Face demo. It offers nine caption types and 17 toggleable instructions. For tag-style captions, the WD14 tagger extension can process images in bulk.

Can I use game sprites for LoRA training? Yes. Spriters Resource and Creative Uncut host sprite sheets and artwork organized by console and franchise. These work well for pixel-art or retro-style LoRAs, though you may need to extract individual sprites.

What is the difference between JoyCaption and GPT-4o for captioning? On a 15-image validation set, JoyCaption Alpha Two matched GPT-4o on average (22.8 vs. 22.7) but showed more variance. GPT-4o was more consistent, while JoyCaption occasionally scored higher on individual images. Source

How do I clean watermarks from training images at scale? Run a free web-based watermark remover first, then pass the output through a denoiser, and finally upscale if needed. Use incognito mode with cookie clearing between sessions to avoid hitting free-tier limits.

Can I mix caption styles in one training dataset? The dataset creator workflow supports toggling between saving images and captions as separate files. However, mixing caption styles within a single training run can confuse the model. Consistent caption formatting across the dataset is recommended.

Sources