# LoRA Dataset Studio > LoRA Dataset Studio (LDS) is a free, self-hosted LoRA training studio. It builds character, style and concept datasets from one reference photo or an image folder, captions them (JoyCaption or a local vision model, prose or booru tags), trains LoRAs with ai-toolkit for Krea 2, FLUX.1, FLUX.2 Klein, Qwen-Image 2.1, Z-Image, SDXL and Anima, and compares checkpoints side by side. It runs on your machine in one browser tab and uses ComfyUI for local generation and checkpoint tests. - License: PolyForm Noncommercial 1.0.0. Noncommercial use is free; commercial use requires separate permission from the licensor. LDS is not MIT-licensed. - Source code, issues and downloads: https://github.com/perfectgf/lora-dataset-studio (Windows ZIP, Git checkout, Docker, Pinokio). - Windows 10/11 runs the full local stack; Linux and macOS work for API-only use and a manual virtual environment; GPU Docker depends on host NVIDIA support. - Training uses ai-toolkit and local generation uses ComfyUI. Importing, sorting and manual captioning need no GPU or API key. - New versions reach Patreon members first, then everyone; the installed version stays free and keeps working. - LDS is not affiliated with loradataset.com. ## Start here - [LoRA Dataset Studio V2](https://perfectgf.github.io/lora-dataset-studio/): Free local LoRA training studio: turn a photo or folder into a captioned dataset, train Krea 2, FLUX, Qwen-Image, Z-Image or SDXL LoRAs, compare checkpoints. - [Documentation](https://perfectgf.github.io/lora-dataset-studio/docs.html): Every LoRA Dataset Studio guide in one place: installation, requirements, the LoRA workflow, dataset quality, settings, troubleshooting and the FAQ. - [Getting started](https://perfectgf.github.io/lora-dataset-studio/guide/getting-started.html): First steps in LoRA Dataset Studio: complete Setup, build a dataset from one reference photo, caption it, train a LoRA and rank its checkpoints. - [Installation](https://perfectgf.github.io/lora-dataset-studio/guide/installation.html): Install LoRA Dataset Studio with the Windows ZIP launcher, a Git checkout, Docker with or without a GPU, or Pinokio, then complete Setup and add plugins. - [Requirements](https://perfectgf.github.io/lora-dataset-studio/guide/requirements.html): What each LoRA Dataset Studio feature needs, from no GPU at all to local LoRA training: VRAM per model family, disk space, Python and operating system. - [End-to-end workflow](https://perfectgf.github.io/lora-dataset-studio/guide/workflow.html): The full LoRA Dataset Studio route, stage by stage: build the dataset, curate and caption it, train with ai-toolkit, test checkpoints and export the LoRA. ## LoRA training FAQ - [LoRA training FAQ](https://perfectgf.github.io/lora-dataset-studio/faq/): Answers to common LoRA training questions: image counts, captions, base models, steps, VRAM, trainers, checkpoint picking, style, concept and video LoRAs. - [How do I train a consistent character LoRA from one or a few photos?](https://perfectgf.github.io/lora-dataset-studio/faq/character-lora-from-one-photo.html): Build a character LoRA dataset from one reference photo: generate about 25 varied shots locally, reject the ones where the face drifts, then train. - [How many images do I need to train a LoRA?](https://perfectgf.github.io/lora-dataset-studio/faq/how-many-images-to-train-a-lora.html): About 25 varied images make a balanced character LoRA. LDS minimums by family: 12 for Z-Image, 15 for Krea 2, FLUX.1 and FLUX.2 Klein, 20 for SDXL. - [Tags or natural language captions? Where does the trigger word go?](https://perfectgf.github.io/lora-dataset-studio/faq/lora-captions-tags-or-natural-language.html): Caption in prose for Krea 2, FLUX, Z-Image and Qwen-Image 2.1, booru tags for SDXL, either for Anima. LDS puts the trigger word first. - [Which base model should I train my LoRA on?](https://perfectgf.github.io/lora-dataset-studio/faq/which-base-model-for-a-lora.html): Pick the LoRA base model before captioning: Krea 2 for realism, FLUX.1 for its ecosystem, FLUX.2 Klein 4B for mid-range GPUs, Z-Image for speed. - [What steps, rank and learning rate should I use?](https://perfectgf.github.io/lora-dataset-studio/faq/lora-steps-rank-learning-rate.html): LDS sets LoRA steps from your dataset: about 120 per image, 1,500 to 3,500, for a character. Default rank is 16 or 32, learning rate 1e-4. - [Which LoRA trainer should I use? Do I need the command line?](https://perfectgf.github.io/lora-dataset-studio/faq/which-lora-trainer.html): LDS trains LoRAs with ai-toolkit and writes its config for you, so runs start from the browser. Exported ZIPs also feed kohya_ss or OneTrainer. - [How much VRAM do I need to train a LoRA?](https://perfectgf.github.io/lora-dataset-studio/faq/lora-training-vram.html): LoRA training VRAM in LDS: 16 GB+ for Z-Image and SDXL, 16-24 GB for FLUX.2 Klein 4B, 24 GB for Krea 2 at 1024 px. Sorting needs no GPU. - [How do I pick the best epoch or checkpoint?](https://perfectgf.github.io/lora-dataset-studio/faq/pick-the-best-lora-checkpoint.html): Compare LoRA checkpoints at a fixed seed in the LDS Test Studio and keep the earliest one that holds the identity. The last checkpoint is rarely best. - [Can I train a style or concept LoRA?](https://perfectgf.github.io/lora-dataset-studio/faq/style-and-concept-loras.html): Yes. LDS has Style and Concept dataset workflows, with prose-only concept captions, 475 × √n concept steps from 2,000 to 12,000 and face masking. - [Can I train a video LoRA (Wan 2.2, LTX-2, MiniMax H3)?](https://perfectgf.github.io/lora-dataset-studio/faq/video-lora-training.html): Video LoRA training in LDS is experimental: the beta Video lane plugin cuts clips to Wan, LTX and MiniMax H3 frame rules and trains with ai-toolkit. - [Do I need to know ComfyUI to use LoRA Dataset Studio?](https://perfectgf.github.io/lora-dataset-studio/faq/do-i-need-comfyui.html): LDS needs a ComfyUI install for local generation and checkpoint tests, not its node editor. Import, captioning and ai-toolkit training work without it. - [How do I upscale or restore images locally?](https://perfectgf.github.io/lora-dataset-studio/faq/upscale-and-restore-images-locally.html): Upscale or restore LoRA dataset images on your own GPU with the SeedVR2 or Klein Improve plugins for LDS. SeedVR2 keeps the look; Klein can shift it. ## Guides - [Building a good LoRA dataset](https://perfectgf.github.io/lora-dataset-studio/dataset-guide.html): What moves the needle in a LoRA dataset: the model family, how many images and which framings, captions and trigger words, settings and checkpoint choice. - [Using the app](https://perfectgf.github.io/lora-dataset-studio/guide/using-the-app.html): Step-by-step instructions for each part of LoRA Dataset Studio: the guided dataset flow, curation, captioning, training, checkpoint testing and export. - [Feature reference](https://perfectgf.github.io/lora-dataset-studio/guide/features.html): Detailed capabilities, examples and limits of LoRA Dataset Studio and its public plugins, including the experimental video and slider workflows. - [LoRA Dataset Studio compared with kohya_ss, OneTrainer, AI Toolkit and FluxGym](https://perfectgf.github.io/lora-dataset-studio/lora-training-tools-compared.html): How LoRA Dataset Studio compares with kohya_ss, OneTrainer, AI Toolkit and FluxGym for LoRA training: models, captioning, checkpoint tests, installs. - [Docker guide](https://perfectgf.github.io/lora-dataset-studio/guide/docker.html): Run LoRA Dataset Studio in Docker on Windows: the two launchers, GPU and API-only setups, an existing or a fresh isolated ComfyUI, updates and restarts. - [RunPod guide](https://perfectgf.github.io/lora-dataset-studio/guide/runpod.html): Run the whole studio on a rented NVIDIA GPU with RunPod and reach it from any browser, with your datasets on a network volume that survives pod restarts. - [Move an existing installation to V2](https://perfectgf.github.io/lora-dataset-studio/guide/migrate-to-v2.html): Switch an existing LoRA Dataset Studio installation from the frozen v1 branch to the maintained v2 branch and keep its datasets and settings. ## Reference - [Settings reference](https://perfectgf.github.io/lora-dataset-studio/guide/settings-reference.html): Every LoRA Dataset Studio setting explained: what it does, its default, when to change it and the traps to avoid, section by section. - [Troubleshooting](https://perfectgf.github.io/lora-dataset-studio/guide/troubleshooting.html): Fixes for the most reported LoRA Dataset Studio problems, listed symptom first, and how to report a new one with a diagnostic report. - [Known limitations](https://perfectgf.github.io/lora-dataset-studio/guide/known-limitations.html): Current boundaries of LoRA Dataset Studio that are not setup failures, so you know what to expect before you build a dataset or start a run. - [Network access and privacy](https://perfectgf.github.io/lora-dataset-studio/guide/network-access.html): Which external services LoRA Dataset Studio contacts, what data they receive, and how to open protected network access from a phone or tablet. - [Extensions guide](https://perfectgf.github.io/lora-dataset-studio/guide/extensions.html): How optional local extensions in backend/extensions load at startup, register their own API routes and mount their own UI in LoRA Dataset Studio. - [Getting help & reporting problems](https://perfectgf.github.io/lora-dataset-studio/guide/getting-help.html): How to get help with LoRA Dataset Studio: generate a diagnostic report, then ask on Discord or open a GitHub issue. - [Responsible use](https://perfectgf.github.io/lora-dataset-studio/responsible-use.html): Consent, privacy, copyright and platform rules for LoRA training with LoRA Dataset Studio, and why a real person's likeness needs their consent.