About this version
in version 4.0
"The final creation: This could be my last workflow. Depends On Feedback"
⚡ WHAT'S NEW IN THIS MAJOR UPDATE ⚡
✅ Custom Fast-Moded Spectrum
└► Powered by MiniMax H3 for ultra-fast sampling! 🚀
✅ New 8-Step LoRA Integrated
└► minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors 🎯
✅High-Speed Performance
└► Faster & optimized CLIP loading time ⏱️
✅ Brand New UI
└► Redesigned sleek layout for a smoother experience ✨
📌 Pro Tip:
Use v3 for maximum facial consistency and detail.
Use v4 if you need the perfect balance of Medium Speed & Consistency.
in version 3.0
🔮 Key Features:
🔥 ✅ 10 Steps Generation Is Now Possible With Same Full Consistency
Just Added Turbo Lora ).
✅ Patch Sol-Attn New Alternative of Sageattention Speed Up The Workflow but in 2nd Generation
✅ MiniMax H3 Mem Eff Sage Attention Patch : To Reduce Peak Vram Usage
in version 2.0
🔮 Key Features:
🔥 Realtime Preview is Now Possible
Just Enable Node Model Preview Override ).☄️Fast Generation 0% Quality Loss : Added Agressive Spectrum & Sage Attention
Added 8 Reference Images (Note 2 or more referance can take extra time)
All Models & Nodes & Tutorial Added (Direct links)
Easy To Use Workflow
Face Consistency is on Ultra Level even on 0.2 Megapixels
Added Low Vram GPU Mathes
Description
"The final creation: This could be my last workflow. Depends On Feedback"
⚡ WHAT'S NEW IN THIS MAJOR UPDATE ⚡
✦ Custom Fast-Moded Spectrum
└► Powered by MiniMax H3 for ultra-fast sampling! 🚀
✦ New 8-Step LoRA Integrated
└► minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors 🎯
✦ High-Speed Performance
└► Faster & optimized CLIP loading time ⏱️
✦ Brand New UI
└► Redesigned sleek layout for a smoother experience ✨
📌 Pro Tip:
Use v3 for maximum facial consistency and detail.
Use v4 if you need the perfect balance of Medium Speed & Consistency.
FAQ
Comments (17)
This WF runs fast and works well. Questions tho, this is a ref2va WF but it uses fl2va diffusion models? Or it can use both model types? Sorry, new to minimax.
i think both you can try it, i have not tried video editing reference but multi reference image to video its working well
@RedditUser9811 Thanks. Yes, both seem to work but give different output.
Yes, you can change te model to REF, but that model its taking like 6x time to create a video in REFV2V mode D:
Runs perfectly and incredibly fast. 2x faster? than the base h3 i2v template. ALL the speedups are meaningful. Setup was incredibly easy. 10+ second videos are easily within reach without ridiculous gen time using this workflow. Running on an RTX 4070ti Super, in Pop!_OS 24.04 LTS.
Just tested it. Surprisingly fast on a 3060 and 32GB of RAM (DDR4). 5 seconds and 0.3MP - 223 seconds. 6 seconds and 0.4MP - 337 seconds.
I have a question. Why so many references? Can I add multiple different photos? Or how does it work?
Yes you can. Each reference can be called out in your prompt. Ie; use <Picture 1> as the background, woman 1 is the person centered in<Picture 2> wearing a dress; as an example. The possibilities are wild for open weight.
First of all, a huge thank you for the truly awesome workflow. Even though I installed SageAttention according to the guide, it still crashed the process. Even with it disabled, the video generation speed at 0.2 took 327.45 seconds (24 fps / 8 seconds). This is amazing!
Please tell me, will there be a workflow version for ITV?
yes all you need to add a node image to video for h3 just add it
for this workflow, do you use --lowvram when starting comfyui?
nope it reduce speed generation time
Thanks a lot for your work. It runs really fast. With my notebook (GeForce RTX 4080 12GByte VRAM and 32GByte RAM) for 10 seconds movie with 0.4 MegaPixels it takes 6,5 minutes. Really the fastest i tested since yet. No problems with Loras for Minimax H3 and lip-synched. Great 😍😍😍😍
Hello. Could you please tell me how to increase the video duration in your workflow?I am trying different H3 workflows and none of them work on my 4070 with 12GB VRAM. Even with the lowest resolution and 2 second duration, this workflow apparently runs out of memory (according to the console window) while loading the MiniMax model and reaching 10500 MB.
All the other stuff like Wan seem to run fine for me, but all the H3 workflows run out of memory no matter how low the settings are.
Also, the Workflow names 2 different models, and the pinkCherry_h3_fl2va_pruned_int8_v0.5-alpha.safetensors one is impossible to find on google, so i used the FL2VA Pruned int 8 one that is mentioned in the text on the left.
Anyone got any ideas why i might be running out of VRAM?
idk, should be fine with 12gb vram,idk how much ram u have or free space,steps...get some 12gb workflow with links for models,maybe problem with comfy,install ComfyUI-Easy-Install