uploadandnap.com
join us ↗
[menu]
home blog lora packs workflows dataset tool faq join us ↗

$ ./dataset-prep --version → v2.0 · multi-api · model-optimized

professional lora dataset prep.

Upload images and get trainer-ready datasets: multi-API captioning and checkpoint-specific optimization for SDXL, Flux, SD-1.5, Qwen and more, with full control over caption quality and style.

see features join skool ↗
dataset-prep --help
  • gemini, openai, grok apis
    Choose the best API for your budget and quality needs.
  • checkpoint-specific prompts
    Auto-optimized captions for SDXL, Flux, SD-1.5, Qwen, WAN.
  • advanced caption controls
    Quality levels, temperature, token limits, quality/token scoring.
  • kohya export ready
    Proper formatting for immediate training use.

$ cat CHANGELOG.md

what makes it work.

Built from the ground up for LoRA training, with intelligence at every step.

new in v2

Multi-API Vision

Switch between Gemini, OpenAI and Grok vision APIs. Optimize for cost (Gemini at $0.14/1K images) or quality. Smart routing gets you the best results for your workflow.

new in v2

Checkpoint Optimization

Model-specific captioning strategies for SDXL, Flux, SD-1.5, Qwen-Image and WAN-2.2. Each checkpoint gets tailored prompts for maximum training effectiveness.

new in v2

Advanced Caption Controls

Fine-tune quality levels, temperature, token limits and caption scoring. Balanced captions for character, style, concept and product LoRAs.

export

Smart Trainer Naming

Files renamed for Kohya format with bucket/repeat awareness. Compatible with all popular trainers. No manual renaming.

output

Rich Text Descriptions

Trainer-friendly .txt files per image. Goal-aligned phrasing, consistent token ordering, ready for immediate training.

debug

Comprehensive Logging

Full transparency into every step: API responses, caption generation and scoring. Debug and optimize with confidence.

$ ls ./use-cases

perfect for.

characterAutomatically avoids captioning "baked features" like hair and eye color, preventing training conflicts.
styleOptimized prompting for artistic styles, visual concepts and aesthetic training.
productDetailed captioning for product design, interior architecture and technical subjects.
fashionPrecise garment descriptions and style metadata for fashion-focused training.
poseAutomatic pose detection and cinematic element tagging for dynamic subjects.
mixedMultiple model types in one dataset, with optimization per checkpoint.

$ cat faq.txt

faq.

Which API should I choose?

Gemini is the most cost-effective at ~$0.14 per 1,000 images. OpenAI is ~$2.16/1K and Grok ~$4.05/1K. The tool recommends an API for your checkpoint type, but you can override it. On a budget, Gemini is unbeatable; for maximum quality, try OpenAI.

Does it work with my checkpoint?

It includes optimized configurations for SDXL, Flux, SD-1.5, Qwen-Image and WAN-2.2, each with model-specific captioning strategies. For other checkpoints, customize the settings or send a feature request.

What training formats does it export?

Kohya format with proper filename patterns for bucket and repeat awareness, compatible with Kohya, AI Toolkit and Civitai Trainer. Each image gets a matching .txt caption file.

Can I control caption quality and style?

Yes. Adjust quality level (standard, detailed, expert), temperature, token limits and quality/token scoring, with different controls per LoRA goal (character, style, concept and more).

Is my data private?

Images are sent to your chosen API (Gemini, OpenAI or Grok) for captioning, under that provider's terms. The tool itself doesn't store or log your images — everything runs in your browser session.

$ ./dataset-prep --start

ready to prep your dataset?

Start building LoRA datasets today.

join skool open tool ↗