# I made a free all-in-one LoRA trainer for consumer GPUs (Windows + Linux): Qwen-Image 2.1, FLUX.2 Klein 9B, Krea 2, Z-Image, Ideogram 4, Anima, SDXL/Pony/Illustrious, LTX 2.3 and MiniMax-H3 (video + audio) — from 4-8 GB VRAM — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/a0dff1ba7a7a830a

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-04 12:51 ingested 2026-10-05 02:39

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1wxey7r/i_made_a_free_allinone_lora_trainer_for_consumer/)

## retained text



```
Hi everyone! I've been working on AcademiaSD LoRAlab Trainer Studio: one installer, one launcher and 9 trainers that share the same web interface. Every model is loaded in 4-bit NF4, the text encoder and VAE run only once in a pre-cache stage, and the whole GPU goes to training.

Platform: made for Windows (one-click installer), with Linux support just added. NVIDIA GPU, RTX 20xx / GTX 16xx or newer.

GitHub: [https://github.com/AcademiaSD/AcademiaSD\_LoRAlab-TrainerStudio](https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio)

WHAT IT TRAINS (minimum VRAM)

\- Qwen-Image 2.1: image LoRAs + edit LoRAs from before/after pairs (8 GB)

\- FLUX.2 Klein 9B: image LoRAs + edit LoRAs (12 GB)

\- Krea 2: image LoRAs, Raw and Turbo (8 GB)

\- Z-Image: image LoRAs, also Z-Image-Turbo (8 GB)

\- Ideogram 4: image LoRAs with JSON captions (12 GB)

\- Anima: anime / illustration LoRAs (4 GB)

\- SDXL: Base, Pony, Illustrious, NoobAI, Juggernaut, RealVis or your own checkpoint (4 GB)

\- LTX 2.3 (2.5): character and style LoRAs for the video model (12 GB)

\- MiniMax-H3: video LoRAs from images, clips and audio, plus RefMods (8 GB)

MINIMAX-H3: VIDEO + AUDIO, AND REFMODS

MiniMax-H3 is a 33B model that generates video and audio together. The official checkpoint is about 500 GB; the trainer uses a 41 GB NF4 version that fits in 8 GB of VRAM with block swap.

\- Datasets can be images, video clips, audio, or any mix.

\- One button prepares your clips (24 fps, valid frame counts) without cutting them.

\- RefMods: encode a few reference images, video clips, or clips with audio into a file that ComfyUI's MiniMaxH3ReferenceToVideo node uses as a native reference. No training: seconds instead of hours.

SHARED FEATURES

\- Automatic captioner (Qwen3-VL) in the dataset manager: natural language, Danbooru tags, or JSON with bounding boxes for Ideogram.

\- Live previews while training.

\- Exact-step resume.

\- One-click export to your ComfyUI / Forge LoRA folder.

\- Remote access from other devices on your network, with a password.

\- SDXL in NF4 trains in about 3.5 GB of VRAM, so 4 GB laptops can train SDXL / Pony / Illustrious LoRAs.

A NOTE

Many models were added almost at the same time, so there may be bugs, or the default settings may not be the best ones. Previews are there to follow the training; judge the final quality in ComfyUI after tuning the LoRA strength. If you find problems or settings that work better, please share them in GitHub Issues. Thanks!

    UPDATE — RunPod support + remote access

    

    Thanks for all the feedback! Based on your comments, these are now in:

    

    - ☁️ RunPod template: one click and it's running in the cloud, no install. It uses a prebuilt Docker image (CUDA 13, Python 3.13, all dependencies), and the code updates itself from GitHub on every start. Pick a GPU, open the address shown in the pod log, log in, and train. A quick SDXL test cost me less than $0.10.

      Deploy on RunPod: https://console.runpod.io/deploy?template=lfxvtg5rrb

    - 🌐 Remote access: use the trainer from another PC, tablet or phone, with a login, or behind your own reverse proxy (external auth and a custom public URL are supported).

    - ⬆⬇ Upload / download in the browser: upload your dataset (files or a .zip, drag and drop works) and download the finished LoRA, handy for remote or cloud setups.

    - 🐍 Your own environment: there's now a requirements.txt, and LORALAB_PYTHON lets you run it from conda/uv instead of the bundled venv.

    

    Remember to Stop and Terminate your pod when you finish: closing the browser doesn't stop billing.

    

  

```



document fetched 2026-10-05 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1wxey7r/i_made_a_free_allinone_lora_trainer_for_consumer
