LoRA training FAQ

These pages answer the questions that come up most in LoRA training threads, with what LoRA Dataset Studio (LDS) does about each; the dataset guide has the full detail.

• How do I train a consistent character LoRA from one or a few photos? Generate about 25 varied shots from one reference photo with Klein or Krea 2 Edit on your ComfyUI, then reject the shots where the face drifts.
• How many images do I need to train a LoRA? About 25 varied images make a balanced character LoRA; the minimum is 12 for Z-Image, 15 for Krea 2, FLUX.1 and FLUX.2 Klein, and 20 for SDXL.
• Tags or natural language captions? Where does the trigger word go? Prose for Krea 2, FLUX.1, FLUX.2 Klein, Z-Image and Qwen-Image 2.1, booru tags for SDXL, either for Anima, with the trigger word first, added by LDS.
• Which base model should I train my LoRA on? Krea 2 has the highest realism ceiling, FLUX.1 the largest LoRA ecosystem, FLUX.2 Klein 4B fits mid-range GPUs, Z-Image iterates fastest, SDXL suits booru checkpoints and Anima targets anime.
• What steps, rank and learning rate should I use? About 120 steps per image for a character, kept between 1,500 and 3,500, with rank 16 or 32 by family and a learning rate of 1e-4.
• Which LoRA trainer should I use? Do I need the command line? LDS drives ai-toolkit and writes its config, so runs start from the browser once ai-toolkit is installed; exported ZIPs also feed other trainers.
• How much VRAM do I need to train a LoRA? 16 GB or more for Z-Image and SDXL, 16 to 24 GB for FLUX.2 Klein 4B and 24 GB for Krea 2 at 1024 px; sorting and manual captioning need no GPU.
• How do I pick the best epoch or checkpoint? Compare checkpoints at a fixed seed in the Test Studio and keep the earliest one that holds the identity.
• Can I train a style or concept LoRA? Yes: Style and Concept datasets have their own captioning rules and step policies, and concepts can mask faces out of the loss.
• Can I train a video LoRA (Wan 2.2, LTX-2, MiniMax H3)? Experimentally, with the beta Video lane plugin; only Wan 2.2 14B has been through a finished run so far.
• Do I need to know ComfyUI to use LoRA Dataset Studio? You need a ComfyUI install for local generation and checkpoint tests, but never its node editor.
• How do I upscale or restore images locally? Use the SeedVR2 plugin to restore with tiling, or Klein Improve to re-render detail, both on your local ComfyUI.