Minimax H3 15-Second Seamless Multi-Shot Full Audio Image-to-Video ComfyUI Template for 12GB GPUs

Minimax H3 15-Second Seamless Multi-Shot Full Audio Image-to-Video ComfyUI Template for 12GB GPUs - v1.1

NSFW

This lightweight workflow is highly optimised so that even lower-end GPUs with only 12GB of VRAM can generate seamlessly stitched 15-second image to video output with fully integrated audio, voices, and sound effects in just 21 minutes.

The template is extremely easy to use, only requiring users to select an image and type in the prompt. That's it. Click run and you're good to go. It's designed to be as user-friendly as the typical cloud image-to-video services.

It's finely optimised to cut as much as possible without falling into high-hallucination range. You can, of course, mess with the resolution output, steps, and turbo settings if you want to speed up generations (at 8 steps you can complete all 15 seconds in just 18 minutes) or go for the full 20 steps at lossless quality, which takes on average, 37 minutes.

This was all made possible thanks to Joey Gambino's H3 Multishot Sampler, some caching prompts, and some specific H3 Minimax optimisers.

Everything you need to run this template is included in the instructions within the JSON.

⚠️ DEPENDENCY REQUIREMENT FOR THE SPECTRUM/SPEED RIG:
This workflow utilizes Patch Sage Attention KJ to push generation speeds to the absolute limit. If you get a red missing node error, click Manager -> Install Missing Custom Nodes to install it.
Note for Windows/Desktop Users: To prevent python terminal crashes, you must manually install the backend math libraries. Open your ComfyUI command prompt/terminal (or embedded python environment) and run:
pip install triton sageattention (or sageattention2 depending on your environment version).
If your system doesn't support SageAttention, you can safely Bypass (Ctrl + B) the Sage Attention node and run the model wire directly through the Spectrum node!

Description

This update based on the original workflow adds a ton of user friendly UX updates, including a much simpler process for generating images to video, using a basic image uploader and text prompt. No more having to fish through the subnodes to find what you need. They're also highlighted in green boxes, so it's super easy to use.

Fixed the audio bug from the first version, so there is continuous audio throughout the whole 15 generation runtime.

Optimised the workflow and cut out unnecessary bloat. Generation times have cut down by 57%, with hardly a loss in quality.

This is highly attuned for lower end machines looking to generate 15-second videos using MiniMax H3 locally.

⚠️ DEPENDENCY REQUIREMENT FOR THE SPECTRUM/SPEED RIG:
This workflow utilizes Patch Sage Attention KJ to push generation speeds to the absolute limit. If you get a red missing node error, click Manager -> Install Missing Custom Nodes to install it.

Note for Windows/Desktop Users: To prevent python terminal crashes, you must manually install the backend math libraries. Open your ComfyUI command prompt/terminal (or embedded python environment) and run:
pip install triton sageattention (or sageattention2 depending on your environment version). If your system doesn't support SageAttention, you can safely Bypass (Ctrl + B) the Sage Attention node and run the model wire directly through the Spectrum node!