Uploaded August 2026 | Updated November 2026, 2 days ago
I made MiniMax H3 almost three times faster in ComfyUI and also tested the new Turbo LoRA, FL2VA, and Ref2VA workflows.
You will learn how to reduce generation time, use low-VRAM models, copy motion from a reference video, combine multiple reference images, and control the first and last frames.
In my speed test, a generation that normally took around four minutes was reduced to around one minute and twenty-five seconds. I used the local Patch Sage Attention KJ node, reduced the sampling steps from 20 to 15, and added EasyCache.
I also tested the MiniMax H3 Turbo LoRA. The four-step setting produced a poor result in my workflow, but eight to ten steps gave a better result. The Turbo LoRA test is separate from the almost three-times speed optimization.
This video covers:
• MiniMax H3 FL2VA and Ref2VA
• Text-to-video and image-to-video
• Low-VRAM and compressed model options
• SageAttention and EasyCache optimization
• Turbo LoRA testing
• First-frame and last-frame control
• Multiple image references
• Character and environment consistency
• Reference-video motion transfer
• Reference audio replacement
• Detailed prompts with dialogue and timing
• Blurry-face fixes and video upscaling
Resources
Workflow: aistudynow.com/minimax-h3-comfyui-workflow-almost-3x-faster-ref2va-guide
MiniMax H3 models:
huggingface.co/Comfy-Org/MiniMax-H3/tree/main
Turbo LoRA:
huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/tree/main
Required nodes:
ComfyUI-KJNodes — Patch Sage Attention KJ and Resize Image v2:
github.com/kijai/ComfyUI-KJNodes
SageAttention Windows installation files:
github.com/woct0rdho/SageAttention/releases
NVIDIA RTX Video Super Resolution:
github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI
ComfyUI official MiniMax H3 guide:
docs.comfy.org/tutorials/video/minimax/minimax-h3
ComfyUI core:
github.com/Comfy-Org/ComfyUI
The Comfy-Org model repository contains the FL2VA, Ref2VA, text-encoder, video-VAE, and audio-VAE files used by native ComfyUI workflows. The official ComfyUI guide identifies KJNodes as the source of Patch Sage Attention KJ
Generation time and memory use may differ based on your graphics card, resolution, model, and workflow settings.
Which MiniMax H3 feature should I test next? Share your graphics card and video memory in the comments.
#MiniMaxH3 #ComfyUI #TurboLoRA
I made MiniMax H3 almost three times faster in ComfyUI and also tested the new Turbo LoRA, FL2VA, and Ref2VA workflows.
You will learn how to reduce generation time, use low-VRAM models, copy motion from a reference video, combine multiple reference images, and control the first and last frames.
In my speed test, a generation that normally took around four minutes was reduced to around one minute and twenty-five seconds. I used the local Patch Sage Attention KJ node, reduced the sampling steps from 20 to 15, and added EasyCache.
I also tested the MiniMax H3 Turbo LoRA. The four-step setting produced a poor result in my workflow, but eight to ten steps gave a better result. The Turbo LoRA test is separate from the almost three-times speed optimization.
This video covers:
• MiniMax H3 FL2VA and Ref2VA
• Text-to-video and image-to-video
• Low-VRAM and compressed model options
• SageAttention and EasyCache optimization
• Turbo LoRA testing
• First-frame and last-frame control
• Multiple image references
• Character and environment consistency
• Reference-video motion transfer
• Reference audio replacement
• Detailed prompts with dialogue and timing
• Blurry-face fixes and video upscaling
Resources
Workflow: aistudynow.com/minimax-h3-comfyui-workflow-almost-3x-faster-ref2va-guide
MiniMax H3 models:
huggingface.co/Comfy-Org/MiniMax-H3/tree/main
Turbo LoRA:
huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/tree/main
Required nodes:
ComfyUI-KJNodes — Patch Sage Attention KJ and Resize Image v2:
github.com/kijai/ComfyUI-KJNodes
SageAttention Windows installation files:
github.com/woct0rdho/SageAttention/releases
NVIDIA RTX Video Super Resolution:
github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI
ComfyUI official MiniMax H3 guide:
docs.comfy.org/tutorials/video/minimax/minimax-h3
ComfyUI core:
github.com/Comfy-Org/ComfyUI
The Comfy-Org model repository contains the FL2VA, Ref2VA, text-encoder, video-VAE, and audio-VAE files used by native ComfyUI workflows. The official ComfyUI guide identifies KJNodes as the source of Patch Sage Attention KJ
Generation time and memory use may differ based on your graphics card, resolution, model, and workflow settings.
Which MiniMax H3 feature should I test next? Share your graphics card and video memory in the comments.
#MiniMaxH3 #ComfyUI #TurboLoRA
![WAN 2.2 Animate in ComfyUI: InfiniteTalk + Uni3C — Low VRAM (GGUF Q4) [Free Workflow]
Download the free workflow: https://aistudynow.com/wan-2-2-animate-comfyui-infinitytalk-uni3c-q4-low-vram/
#ComfyUI #WAN22 #InfinityTalk
Swap any face or object inside one ComfyUI flow. Blend looks real. Lips match. Camera move support with Uni3C. FP8 vs GGUF Q4 test on low VRAM.
what you will learn
InfiniteTalk with WAN 2.2 Animate for better lip sync
Uni3C camera motion test and when to keep it off
FP8 vs Q4 GGUF on low VRAM and what to pick
LoRA we used
Relighting LoRA for Animate to match scene color and tone WanAnimate_relight_lora_fp16.safetensors
LightX 4-step LoRA set for Animate or I2V
low_noise_model.safetensors
Pusa v1 motion LoRA if you want stronger motion : Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors
HPS LoRA for human-preference quality tuning : Wan2.2-Fun-A14B-InP-LOW-HPS2.1_resized_dynamic_avg_rank_15_bf16.safetensors
upload the edit video and pick a supported WAN 2.2 size with Resolution Master
upload your subject image at the same size, remove background
mask with points: green keep, red remove, test mask once
models:
Wan2.2-Animate-14B FP8 model: Wan2_2-Animate-14B_fp8_e4m3fn_scaled_KJ.safetensors
Folder: ComfyUI/models/diffusion_models
Optional for low VRAM: Wan2.2 I2V GGUF Q4 or Q5 pairs here is the link
disclosure and credits
this video uses AI to alter real footage for an educational demo. altered or synthetic content label is on in Studio.
WAN, InfinityTalk, Uni3C, LoRA models belong to their creators. credits to them. WAN 2.2 Animate in ComfyUI: InfiniteTalk + Uni3C — Low VRAM (GGUF Q4) [Free Workflow]](https://i.ytimg.com/vi/fB7rCasr6yQ/mqdefault.jpg)









