AcademiaSD LoRAlab: Nine LoRA Trainers Behind One Launcher

The launcher: pick a trainer and it opens in the browser, while models, projects and settings stay shared.

What it trains, and on how much VRAM

The floor is an NVIDIA GPU with compute capability 7.5 or newer (RTX 20xx / GTX 16xx and up), a 580-series driver or newer for CUDA 13, and 16 GB of system RAM. MiniMax-H3 video clips additionally need ffmpeg on the PATH. GTX 10xx cards are out: PyTorch for CUDA 13 starts at the RTX 20xx generation.

One interface per trainer

The Qwen-Image 2.1 trainer. All nine trainers share the same panels.

Every trainer follows the same five steps:

1. Project and dataset. Name the project, add an optional trigger word, point at a dataset folder. Each image needs a .txt caption with the same name, and images can be dragged in as a folder or a .zip.
2. Captions (optional). A vision-language model writes a caption for every image with the trigger word first; the prompt is editable and existing captions are kept unless you overwrite them.
3. Pre-cache. The text encoder and the VAE run once at your chosen resolution and their outputs are stored on disk. Re-running skips images that did not change, and afterwards 100% of the GPU goes to training.
4. Train. Set steps, learning rate, rank and alpha, then start. Settings can be changed mid-run and apply on the next step, and Stop Training saves the exact step, so a run resumes days later without losing a step.
5. Export. Send to Models copies the LoRA into a ComfyUI (or Forge / A1111) models/loras folder that the studio remembers across trainers. Step checkpoints can also be downloaded from the browser, and exported LoRAs carry kohya-style metadata that Civitai and LoRA managers read.

The studio works from another machine on the same network too: the launcher has a remote access mode with a username and password, so a dataset can be uploaded from a laptop and the trained LoRA downloaded back without touching the training PC.

Sharp previews while training

Some of these models are slow enough that sampling needs to be watched, and ComfyUI's default Latent2RGB preview for Qwen-Image 2.1, Krea 2 and FLUX.2 is blocky color. The README's recommended pair is kijai's KJNodes plus AcademiaSD's tiny autoencoders in ComfyUI/models/vae_approx/: TAE-Qwen-Image-2.1, TAE-Krea-2 and TAE_Flux2_AcademiaSD, dropped into the Model Preview Override node between the model and the sampler.

Install

One installer per platform, and one shared venv for every trainer (PyTorch 2.14 on CUDA 13.0, Diffusers from GitHub, Transformers, PEFT, bitsandbytes and Flask):

git clone https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio.git

On Windows, double-click Install_LoRAlab-TrainerStudio.bat; on Linux every .bat has a .sh equivalent, and the launcher picks the right one. Install_Triton&SageAtten220.bat adds Triton and SageAttention 2.2 if you want them. Updating is a single Update_LoRAlab-TrainerStudio.bat, which keeps models, datasets, projects and settings. The project also ships a RunPod template and a Docker image for cloud training.

The MiniMax-H3 trainer, one of the two video trainers in the studio.

The author flags the studio as early work: the trainers were added in quick succession, so default training and preview settings may not suit every model, and issues are the intended way to report what breaks.

Availability

AcademiaSD LoRAlab Trainer Studio lives at github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio, with pre-quantized training packages such as AcademiaSD/Qwen-Image-2.1-NF4-for-LoRA-Training on Hugging Face.