LoRA Dataset Studio compared with kohya_ss, OneTrainer, AI Toolkit and FluxGym

kohya_ss and OneTrainer fit people who want a dedicated trainer with many settings and model families, and FluxGym fits a FLUX.1 LoRA with the fewest choices on a 12 to 20 GB card. AI Toolkit has the longest model list of the four trainers, including Krea 2, Qwen-Image 2.1, FLUX.2 and video models, and it is the engine LoRA Dataset Studio (LDS) trains with. LDS adds the work around each run for seven image families, among them Krea 2, FLUX.1, FLUX.2 Klein and Qwen-Image 2.1: building, generating and captioning the dataset, launching guided runs and comparing checkpoints at a fixed seed through ComfyUI. Choose LDS when the dataset or the choice of checkpoint is the hard part, and a trainer on its own when you want full control of the training configuration or a model family LDS has no recipe for.

Checked against each project's README in November 2026.

Not documented means the README does not say; the feature may still exist elsewhere in that project. LDS entries come from the LDS documentation.

Yes means the family is named in that project's README; for LDS, in its documentation.

• You start from one reference photo: LDS proposes the face, bust, body and back shots a character set needs, generates the missing ones locally with Klein or Krea 2 Edit, and face similarity helps you reject shots where the face drifts.
• You start from a large folder: the Image Bank scores images, groups duplicates and people, and builds balanced shortlists to promote into a dataset.
• You want captions in the style each family expects (prose for Krea 2, FLUX and Z-Image, booru tags for SDXL), with the trigger word added on export and per-trait rules for what the captions describe.
• You would rather pick a recipe than write a config: LDS writes the AI Toolkit config for its seven image families, scales steps to the dataset, queues runs and continues from any saved epoch.
• You choose checkpoints by looking: Test Studio renders the checkpoints you pick at several strengths with the same seed and prompts, then lets you vote and rank them.

Datasets export as standard ZIPs of images with same-name .txt captions, so a set built in LDS can feed kohya_ss, OneTrainer or another trainer, and captions written in another tool can be imported back.

• You want a GUI over a wide range of sd-scripts parameters, with the CLI command generated for each run.
• You need LoHa, Dreambooth, Textual Inversion or LECO concept erasure, which LDS does not offer, or fine-tuning beyond the experimental Krea 2 full-model lane in LDS.
• You train SD 1.5, SD 2.x, SD3, Lumina Image 2.0 or HunyuanImage-2.1, which have no LDS recipe.
• You train locally on Linux, which its README supports through community-maintained instructions, while the full local LDS stack targets Windows 10/11. Its README also documents a GUI that works offline once installed.

• You want full fine-tuning or embeddings as well as LoRA; LDS fine-tunes only Krea 2, experimentally and in the cloud.
• You train families LDS has no recipe for, such as Chroma, SD 3.x, Sana, PixArt, Stable Cascade, Ernie Image or Hunyuan Video.
• You want the training options its README lists in one place: masked training, EMA, aspect-ratio bucketing, multi-resolution training, noise-schedule rescaling, image augmentation, several prompts per image and automatic backups to continue a run.
• You want BLIP, BLIP2 or WD-1.4 captioning, mask generation and model format conversion built in, with install scripts for Windows, Linux and macOS.

AI Toolkit is the engine LDS drives, so this is a choice of interface, not of trainer.

• You need a model LDS has no recipe for. Its README lists over 30 image models, edit models such as Qwen-Image-Edit and FLUX.1 Kontext, Wan, LTX and MiniMax-H3 video models, and audio models; the LDS feature reference itself recommends AI Toolkit directly for unsupported architectures and experimental configuration keys.
• You want to edit the YAML config yourself, train only chosen layers or switch the network to LoKr.
• You run Linux (including ARM64), macOS on Apple Silicon or a headless server, where the experimental AI Toolkit Manager keeps its whole environment inside its own folder.
• Your dataset is already built and captioned and you only need the run.

LDS uses an existing AI Toolkit folder set in Settings → Local tools, so one install can serve both.

• You want a FLUX.1 LoRA in three steps: enter the LoRA details, upload and caption the images, start.
• Your card has 12, 16 or 20 GB of VRAM, the sizes its README targets.
• You prefer a Pinokio one-click or Docker install, with models downloaded when training starts.
• You want sample images every N steps, publishing to Hugging Face from the same UI, and an Advanced tab with every sd-scripts flag when the defaults are not enough.

Its README lists FLUX.1 bases only; the family table above shows which tools name Krea 2, FLUX.2 or Qwen-Image.

• Training needs AI Toolkit. Local training uses a separate AI Toolkit install, set in Settings → Local tools; Qwen-Image 2.1 needs a recent toolkit with qwen_image_2 support. LDS has recipes for seven image families, not for every model AI Toolkit supports.
• Generation and testing need ComfyUI. Generating dataset shots locally, Test Studio and Canvas generation need a reachable ComfyUI with matching models; Setup downloads the generation models. Training on official Hugging Face bases does not need ComfyUI, but SDXL training takes its base checkpoint from ComfyUI's model folders.
• VRAM. Importing, sorting and manual captioning need no GPU. Auto-captioning takes about 8 GB of VRAM and local generation about 16 GB, on NVIDIA cards; training needs the amounts in the table above, and Qwen-Image 2.1 has not been measured in LDS. 16 GB of system RAM is recommended, and the app does not check it.
• Platform. The full local stack targets Windows 10/11; Linux and macOS run the API-only mode with a manual venv. The Docker images and the RunPod image do not include AI Toolkit: the GPU Docker lanes use a host install or cloud training, and the no-GPU image and the RunPod pod have no local training.
• Experimental parts. Video LoRA training is beta, and only Wan 2.2 14B has completed a run so far; slider LoRAs are beta; full-model Krea 2 training is experimental, cloud-only and needs 80 GB of VRAM. Cloud training support varies by family, and FLUX.1 and Anima train locally only.
• License and releases. PolyForm Noncommercial 1.0.0: commercial use needs separate permission from the licensor. New versions reach Patreon members first, then everyone; exclusive plugins stay members-only, and the installed version keeps working.

The known limitations page lists the current boundaries in detail.

• Requirements: hardware, VRAM and dependencies by feature
• Pick your model family first
• Which LoRA trainer should I use?
• How to pick the best LoRA checkpoint