Minimax H3: Pixel art short video lora training locally

This workflow demonstrates how to train a clip lora for Minimax-H3 with sample pixel art dataset.

Requirements

• Nvidia GPU: 16GB+ VRAM
• Inline Studio installed

Steps

This is a step by step guide to run this workflow.

Step 1: Setup Inline Studio

• Visit installation instruction on our Github readme to setup Inline Studio on local machine with your own GPU.
• No GPU? Launch Inline Studio via official template on Runpod

Step 2: Download workflow

You can download this workflow from the download icon next to title. Click on it & the json file will be saved to your local machine.

Step 3: Drop the workflow

• Open Inline Studio canvas, create a new project or use an existing project
• Drag & drop the saved workflow into the canvas

Step 4: Install models

As soon as you dropped the workflow, missing model popup will guide you through model download.

Models required:

core/models/
  diffusion_models/  minimax_h3_fl2va_bf16.safetensors        <-- raw bf16 for best quality
  text_encoders/     MiniMax-H3-processor/                    <-- folder
  text_encoders/     MiniMax-H3-text-encoder/                 <-- folder     
  vae/               minimax_h3_audio_vae_fp32.safetensors    <-- audio vae  
  vae/               minimax_h3_video_vae_fp16.safetensors    <-- video vae  

Note: Links are given under model section.

Step 5: Load dataset

On the side panel, do to dataset tab, Create a new dataset by adding a name & trigger word: Dataset name: pixel-art Trigger word: inline-pixel

Import dataset from Huggingface Once empty dataset is created, click Add/Manage Training Data & switch to Huggingface tab & paster repo id trojblue/test-HunyuanVideo-pixelart-videos

Hit check to verify & then import all to add all clips & captions into your newly created dataset.

Step 6: Training Params

Click on adjust icon on the bottom of Train Lora node.

Important LoRA parameters

Label	Key	Default	Description
Output LoRA name	outputName	(auto)	Filename for the saved adapter
Max clip length (seconds)	clipSeconds	1	Ceiling on trained clip length(smaller clips are auto included)
Clip window	clipWindow	start	Which end of the clip(set between 1-5s)
Steps	steps	1500	Total training iterations
Resolution	resolution	1024	Training pixel size; dominates VRAM
Save every	saveEvery	250	Checkpoint interval, in steps
Keep a LoRA at every checkpoint	saveSnapshots	false	Usable adapter per save point

Step 7: Start Training

Hit start training after selecting Train lora node & monitor the progress.

Step 8: Monitor Train

You can monitor progress directly via Train Lora node, If you have enabled Save every & Keep a LoRA at every checkpoint, mid training checkpoints will be saved automatically under core/models/loras, you can use them before even the full training finishes.

Step 9: Generated with Lora

If you just want to generate using same generated lora, Setup a new separate generation graph:

Load Lora(Selected your lora) -> Minimax H3 Text to video Generation Node 

For any issue or support, reach me out on Discord.