Home Creators Posts Import Register Favorites Logout

Downloads

Content

Hey everyone! I’ve just released a brand new video on the MINIMAX H3 IMAGE LORA TRAINING! After crowning MiniMax H3 Image the high-resolution editing king in my last video, we’re now making it even MORE powerful by training our own characters, editing effects and art styles inside FIZGIG, an amazing training UI!

You can check out the video right here: https://youtu.be/4xoQxStgFgw

LINK TO THE XY PLOT LORA WORKFLOWS: https://www.patreon.com/aitrepreneur/posts/176002181



LOCAL INSTALL!!! RECOMMENDED TO HAVE AT LEAST 24GB OF VRAM AND 64GB OF RAM FOR THESE PRESETS! If your hardware is limited, use the RunPod installer instead, as shown in the video!

Local install:

Requirements:

– Windows 10/11 64-bit
– A supported NVIDIA GPU or compatible AMD Radeon GPU
– Updated GPU drivers. NVIDIA users need driver 580 or newer
– At least 100GB of free disk space, preferably 150GB or more for your datasets and trained LoRAs

Python, Git, the required dependencies, H3 models and my three presets are handled automatically by the installer!

  1. Download MINIMAX_H3_FIZGIG_INSTALL_WINDOWS.bat and place it in a permanent folder, such as C:\AI. Don’t run it from inside a ZIP or OneDrive.

  2. Double-click the BAT file and select 1 for NVIDIA or 2 for AMD, depending on your GPU.

  3. Wait for the installation and model downloads to finish.

  4. ???

  5. Profit 😎

Fizgig opens at http://127.0.0.1:8081. Keep the console window open while using it.

Once you’ve closed Fizgig and want to start it again, use START_FIZGIG.bat inside the Fizgig-H3-Installer folder. If you placed the installer in C:\AI, the launcher will be here:

C:\AI\Fizgig-H3-Installer\START_FIZGIG.bat

RUNPOD. HERE'S A VIDEO ON HOW TO DO IT: https://youtu.be/mGaubj2guR8

  1. Create an account if you haven’t already: RunPod

  2. Click Pods, then Deploy. Choose a GPU with 24GB or more VRAM, and select the RunPod PyTorch 2.8.0 template with JupyterLab.

  3. Edit the template and set the Volume Disk to 200GB. Keep its mount path as /workspace. Under Expose HTTP Ports, enter:

    8888,8082,8083

  4. Deploy the pod, wait for it to finish starting, then click Connect → HTTP Service on port 8888 to open JupyterLab.

  5. Open the /workspace folder in the file browser on the left and upload MINIMAX_H3_FIZGIG_INSTALL_RUNPOD.sh there. Then open a Terminal.

  6. Copy and paste this command, then press Enter:

    bash "MINIMAX_H3_FIZGIG_INSTALL_RUNPOD.sh"

  7. Wait for the installation and model downloads to finish. The installer also configures the model paths and adds all three training presets automatically.

  8. When the terminal says “Fizgig is ready”, go back to your pod and open Connect → HTTP Service on port 8082. Enter the password printed in the terminal.

  9. To upload datasets or download your trained LoRAs, open port 8083 for the file manager. The username is admin; use the password printed for the file manager.

Save the password shown during installation! If the installer prints different ports because the defaults are occupied, expose those HTTP ports in your pod settings and use them instead.

After stopping and restarting your pod:

Open the JupyterLab terminal again and run:

bash "MINIMAX_H3_FIZGIG_INSTALL_RUNPOD.sh"

This starts both Fizgig and the file manager again. Then reopen their HTTP ports from Connect.

TRAINING PRESETS:

The installer automatically adds:

– H3-PERSON: for a person or character
– H3-EDIT: for an editing effect
– H3-STYLE: for an art style

In Fizgig, select MiniMax H3 Image, choose the appropriate preset, select your dataset and give your training run its own output name. Then follow the training tutorial!

The three JSON presets are also attached separately below.

As always, supporting me on Patreon allows me to keep creating helpful resources like this for the community. Thank you for your support, now go have some fun 😉!

Files

Comments

tyrelli doomah

I have a 5090. What are typical run times in training an H3-Person Lora? Is that something I would want to rent a RUNPOD for? As I recall in a past video it took around 24 hours to train a Lora months ago? With Fizgig, what type of training times are we talking about. Better to save the local 5090 graphics card and use a Runpod?, or is the training time short enough to train locally? Love the videos and json, thanks.

Aitrepreneur

I think for around 20 epoch it's like 1h of training, depends on the gpu of course but 5090 will be really fast, so don't worry about it, you don't need 32gb to train a lora for it so you have wayyy more than enough vram for full speed training

tyrelli doomah

Perfect. I'll try the local run. 1 hour isn't too much time at all. Thanks.

Rob MR

Rip my 4070

Alexandru C.

do you consider a 5080 good enough (efficient as in time consumming vs price/hr of equivalent good GPU online) for locall install and 32 GB of RAM (with enough free space of SSD) or recommend runpod? Also, do you recommend FIZGIG for other models such as MOSS, Krea or QWEN ?