Lonecat's MiniMax H3 FL2VA/ REF2VA w/ Low VRAM Optimizers, Upscaler, & Post Production

MiniMax H3

I just updated my node pack. If it gives you errors, download it from Github (Link Below)

This is an advanced workflow with numerous options. It is not necessarily designed for beginners (although I encourage you to try it). For a simpler version, try my one, Here: https://civitai.red/models/2600919/lonecats-simple-workflows-minimax-h3-krea-2-zit-klein9b-illustrious-flux-wan22-qwen-image-edit

Instagram: https://www.instagram.com/synth.studio.models/

Buy me a☕ https://ko-fi.com/lonecatone

This represents hundreds of hours of work that I give away for free. If you enjoy it, Please make sure to 👍like and post to my workflows. That is how you support me!!! Also, feel free to ⚡tip 😉

Uses my Repo https://github.com/lonecatone23

You need these node packs, or download them from the manager:

ComfyUI_LC123_nodes

LC_AV_nodes

也有中文说明

V6.1

• apparantly had a wire crossed and forgot to hook up the prompt on Ref2VA after messing with a new LLM. Both are repaired.

V6.0

• Added a bunch of new audio and video nodes including a load audio node you can trim in real time:

• 

• Added a optimized VAE that allows you to choose between regular and tiled to help with OOM issues:

• Added an upscale suite that actually works at 1.0 for low VRAM quality🤔

• 

• Added a whole audio group and expanded the video post processing for killer sound and image control

• 

• Updated the save nodes. This is beta. I'm trying to get them to transfer data over to CivitAI.

• Updated and added notes.

V5.0

I went back to the drawingboard. I borrowed an idea (or two) from @Plaguekind, researched several other optimization ideas, and made a change to the LLM:

• Swapped out cheesy LLM prompter for QwenVL with Minimax 🎞️ options.

• Changed the entire setup on how the models swap.

• Moved stuff around to confuse anyone who just got used to the other workflow.

  • Eliminated duplicate loaders.

  • eliminated duplicate Preview windows.

• Swapped out the limited LoRA loaders for a Power LoRA loader.

• Improved the Sampler setup.

• Added and updated notes.

• drank a lot of ☕

V4.0

I spent the last few weeks going through everyone else's workflows to see what worked and what did not. I built some custom nodes to deal with issues and tested out pretty much every optimization setup listed in the ComfyUI manager and on CivitAI

I can get faster, but with less quality. I can go slower, but not gain much more quality. I tried chunking, Sparse attention, Optimizers, Cache options, etc.

I trimmed it down to this:

• Overhauled the reference area

• created pipes to better transfer data

• Added Advanced Sparse Attention and H3 Optimizer nodes.

  • https://github.com/Zironic/H3-Optimizations

    • Speed is significantly faster with better quality than other setups.

    • even works with Sage Attention!!

• Added h3-eros-max capabilities

V3.2

Man, I'm batting 1.000 as of late 🤔

• Added a second LoRA loader

• Replaced the Sage attention nodes with the new H3 ones

• Fixed the break there at the same time 🙄

• made a connection redo in the reference area.

V2.0

• Added Turbo LoRAs

• Added H3 Sigma schedulers

• Massive update in notes

• Added LoRA Loader that works for both models

• Upgraded the LLM prompt Instructions to better align with the Model

Overview

This workflow includes a comprehensive set of features designed for advanced video generation and enhancement:

• Support for GGUF or standard diffusion models and CLIP models

• Integrated LLM for intelligent prompt engineering

• First-frame / last-frame to video (FLF2V) generation

• Reference-to-video (REF2V) capability

• Supports up to 4 reference images (expandable via additional connections)

• Accepts video inputs

• Accepts audio inputs

• RIFE frame interpolation

• RTX Super Resolution upscaling

• 也有中文说明