A workflow uses the latest Open Source MiniMax H3 to generate videos from text.
Other model details are within the workflow.
The Low VRAM version uses the ComfyUI Uniblockcache to Run the MiniMax H3 with 6-8 GB of VRAM.
Try the workflow online(Faster):
Basic: https://www.runninghub.ai/post/2084115262681878529
LowVRAM: https://www.runninghub.ai/post/208570060333119078
LatentUpscaler: https://www.runninghub.ai/post/2089604110974033921
Please check the video tutorial before you use this workflow:
Basic Usage:
https://rumble.com/v7dotpw-seedance-is-dead-minimax-h3-rebornfree-open-source.html
Low VRAM Usage:
https://rumble.com/v7dvf7e-minimax-h3-advanced-usagepromptlow-vramspeed-up-lora.html
Latent Upscaler+Local Prompt:
https://rumble.com/v7een90-minimax-h3-latent-upscaler-local-prompt-skill-nodeqwen3.8-gguf.html
Custom Nodes Used In Workflow:
Custom Nodes:
https://github.com/T8mars/comfyui-minimax-h3-audio-T8
https://github.com/ptmaster/ComfyUI-PT_H3ConcatAVLatent
https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
https://github.com/smthemex/ComfyUI_UniBlockSwap
https://github.com/chflame163/ComfyUI_Qwen_H3_Prompt
https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
Description
This version uses a latent upscaler in the second-stage sampling, saving the time of repeated decoding and encoding. Fix artifacts in the output video. Plus a fully local prompt skill node(Qwen3.8 27b).