RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2086411824722956290?inviteCode=rh-v1111
MiniMax H3 Larry + Sageattention Acceleration — MiniMax H3 Larry + SageAttention Acceleration
A MiniMax H3 video graph from episode 291, part of the six-times-acceleration series that compares model tier, LoRA, attention and cache settings.
Saved configuration
91 nodes / 111 links. Four parallel branches: UNET minimax_h3_fl2va_int8_convrot.safetensors, CLIP qwen3vl_32b_minimax_h3_int8_convrot, video fp16/audio fp32 VAEs, turbo-v4 LoRA at 1, euler + beta 6 steps denoise 1; four LoadImage inputs and four outputs.
How to use
This is the Larry + SageAttention test file: four branches share the same int8 model and turbo LoRA but differ in the attention/cache patches wired around them. It exists to compare speed, not to be the cleanest single pipeline; if you want one graph, use the Larry Turbo file instead.
Limits and not-verified items
No inference was run for this listing: no speed, VRAM, quality, audio-sync or acceleration-factor claim is made. The "6x" in the episode title is the author's description and was not independently measured. Values are quoted from the saved nodes. H3 weights, LoRAs, encoder and VAEs are not bundled; check their licenses separately. The graph notes credit the official Comfy-Org/MiniMax-H3 model files; preserved as-is, no originality claim. The cover image is a promotional still, not a frame produced by this graph.
Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2086411824722956290?inviteCode=rh-v1111
MiniMax H3 Larry + Sageattention Acceleration — MiniMax H3 Larry + SageAttention AccelerationA MiniMax H3 video graph from episode 291, part of the six-times-acceleration series that compares model tier, LoRA, attention and cache settings.
Saved configuration91 nodes / 111 links. Four parallel branches: UNET minimax_h3_fl2va_int8_convrot.safetensors, CLIP qwen3vl_32b_minimax_h3_int8_convrot, video fp16/audio fp32 VAEs, turbo-v4 LoRA at 1, euler + beta 6 steps denoise 1; four LoadImage inputs and four outputs.
This is the Larry + SageAttention test file: four branches share the same int8 model and turbo LoRA but differ in the attention/cache patches wired around them. It exists to compare speed, not to be the cleanest single pipeline; if you want one graph, use the Larry Turbo file instead.
Limits and not-verified itemsNo inference was run for this listing: no speed, VRAM, quality, audio-sync or acceleration-factor claim is made. The "6x" in the episode title is the author's description and was not independently measured. Values are quoted from the saved nodes. H3 weights, LoRAs, encoder and VAEs are not bundled; check their licenses separately. The graph notes credit the official Comfy-Org/MiniMax-H3 model files; preserved as-is, no originality claim. The cover image is a promotional still, not a frame produced by this graph.
Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.