RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2086411704518406146?inviteCode=rh-v1111
MiniMax H3 Int8 W4A8 Acceleration Comparison — MiniMax H3 INT8 W4A8 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
62 nodes / 78 links. Three parallel branches saved in one file: W4A8 mixed UNET minimax_h3_fl2va_pruned_w4a8_mixed.safetensors, plus two minimax_h3_fl2va_int8_convrot.safetensors branches; CLIP qwen3vl_32b_minimax_h3_nvfp4_awq / ..._int8_convrot; int8 and fp32 VAEs. No LoRA nodes. Sampling: res_multistep + simple, 20 steps, denoise 1, three LoadImage inputs and three outputs.
How to use
This is the acceleration comparison file rather than a single production graph: each branch is a different quantization (W4A8 mixed vs int8 conv-rot) and all three save to the same prefix. Read which branch you are queueing before you compare timings — and note this file keeps the full 20-step recipe, unlike the turbo-LoRA graphs.
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/2086411704518406146?inviteCode=rh-v1111
MiniMax H3 Int8 W4A8 Acceleration Comparison — MiniMax H3 INT8 W4A8 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 configuration62 nodes / 78 links. Three parallel branches saved in one file: W4A8 mixed UNET minimax_h3_fl2va_pruned_w4a8_mixed.safetensors, plus two minimax_h3_fl2va_int8_convrot.safetensors branches; CLIP qwen3vl_32b_minimax_h3_nvfp4_awq / ..._int8_convrot; int8 and fp32 VAEs. No LoRA nodes. Sampling: res_multistep + simple, 20 steps, denoise 1, three LoadImage inputs and three outputs.
This is the acceleration comparison file rather than a single production graph: each branch is a different quantization (W4A8 mixed vs int8 conv-rot) and all three save to the same prefix. Read which branch you are queueing before you compare timings — and note this file keeps the full 20-step recipe, unlike the turbo-LoRA graphs.
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.