Skip to content
Agientry
← Back to Videos

Ref2VA-VSA Cuts MiniMax-H3 Generation to 72s

Loading video
Loading video

A ComfyUI node combines sparse attention with dense reference tokens, cutting MiniMax-H3 character video generation from 650s to 72s while using about 13.5GB VRAM.

ComfyUIMiniMax H3Ref2VAVSASparse AttentionDiffusion TransformerVideo GenerationInference Optimization
Category: video-generation
Author: @Forhanvv
Date: 2026-10-16T00:00:00
Duration: 10.333s
Source: https://x.com/Forhanvv/status/2111192898171818232
Reference: https://github.com/Kablex/ComfyUI-Ref2VA-VSA

Related Videos

Sol-H3 Brings MiniMax-H3 to ComfyUI
0:56

Sol-H3 Brings MiniMax-H3 to ComfyUI

@xieenze_jr · 21 views · 2026-10-16
MiniMaxMiniMax H3Sol Engine
LightVAE Cuts MiniMax H3 Decode Time
0:10

LightVAE Cuts MiniMax H3 Decode Time

@ITOKtw · 55 views · 2026-10-15
MiniMax H3LightVAEVideo VAE
How to build cinematic transitions in Google Flow with Omni 1.1 Flash
1:15

How to build cinematic transitions in Google Flow with Omni 1.1 Flash

@FlowbyGoogle · 51 views · 2026-10-15
Google FlowOmni 1.1 FlashVideo Generation
Kling 4.0 launches with native 30 second video generation
3:22

Kling 4.0 launches with native 30 second video generation

@xiaohu · 45 views · 2026-10-15
Kling 4.0KuaishouVideo Generation
BeefTV, an open source desktop studio for local AI short-drama production
0:34

BeefTV, an open source desktop studio for local AI short-drama production

@laowangbabababa · 40 views · 2026-10-15
BeefTVVideo GenerationStoryboard
OmniHuman 2.0: Audio-Driven Avatar Video at Five Minutes
0:58

OmniHuman 2.0: Audio-Driven Avatar Video at Five Minutes

@BytePlusGlobal · 49 views · 2026-10-15
OmniHuman 2.0BytePlusAudio-Driven Avatars

As an Amazon Associate, we earn from qualifying purchases.