https://youtu.be/SUJkkT6XLHI 
 
This workflow is a MiniMax H3 acceleration test built around INT8 and W4A8 quantized routes. It is useful when you want to compare lighter MiniMax H3 video generation branches without rebuilding the full graph from scratch. 
 
The connected graph uses MiniMax H3 FL2VA model routes, Qwen3-VL MiniMax text encoders, MiniMax H3 video VAE, and the MiniMax H3 audio VAE. The layout keeps multiple reachable output branches so you can compare the W4A8 mixed model route against INT8 ConvRot routes in one project. 
 
This package is meant for practical ComfyUI and RunningHub users who want a ready-made MiniMax H3 graph instead of rebuilding loader, encoder, VAE, sampler, and video export connections by hand. The workflow is especially useful for speed comparison, hardware planning, and repeatable prompt tests. Keep the same source image and prompt when comparing versions so the difference you see comes from the workflow route, not from changing creative variables. 
 
Main features: 
 
- MiniMax H3 acceleration-oriented comparison workflow 
- W4A8 mixed and INT8 ConvRot model branches are staged in one graph 
- Qwen3-VL MiniMax encoders are connected to the active generation paths 
- Video VAE and audio VAE routes are included for FL2VA style output 
- Portrait 9:16 resolution selector is used in the connected branches 
- Res-multistep sampling with a simple 20-step schedule appears in the tested branches 
- Designed for quick side-by-side testing of quantized MiniMax H3 behavior 
 
Suggested workflow: 
 
Start by testing one branch at a time and keep the same image, prompt, seed strategy, and duration when comparing model variants. For fair speed and quality checks, avoid changing the prompt between branches, then review face stability, motion consistency, lip movement, and background drift before deciding which branch fits your hardware. 
 
For better comparisons, make one small test first, review identity stability and temporal motion, then raise prompt complexity only after the base route behaves correctly. When using speaking-character prompts, keep the face visible, avoid too many fast cuts, and separate action, expression, and camera instructions clearly. 
 
RunningHub Workflow 
 
Try the workflow online right now - no installation required. 
Workflow: https://www.runninghub.ai/post/2086424109604515841?inviteCode=rh-v1111 
 
If the results meet your expectations, you can later deploy it locally for customization. 
 
Fan Benefits: Register to get 1000 points + daily login 100 points - enjoy 4090 performance and 48 GB super power! 
 
Bilibili Updates (Mainland China &amp; Asia-Pacific) 
 
If you're in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown. 
Bilibili Video: https://www.bilibili.com/video/BV17nuU6xE3g/ 
 
Support Me on Ko-fi 
 
If you find my content helpful and want to support future creations, you can buy me a coffee. 
Every bit of support helps me keep creating. 
Ko-fi: https://ko-fi.com/aiksk 
 
Business Contact 
 
For collaboration or inquiries, please contact aiksk95 on WeChat. 
 
打开下方链接即可在线体验，无需安装。 
工作流：https://www.runninghub.ai/post/2086424109604515841?inviteCode=rh-v1111 
 
如果你觉得效果理想，也可以在本地进行自定义部署。 
 
粉丝福利：注册即可领取 1000 积分，每日登录再领 100 积分，体验 4090 和 48GB 大显存性能。 
 
B站视频（中国大陆及亚太地区） 
 
如果你在中国大陆或亚太地区，可以通过下面的视频查看工作流的实测效果与创作思路。 
B站视频：https://www.bilibili.com/video/BV17nuU6xE3g/ 
 
我会在夸克网盘持续更新模型资源： 
https://pan.quark.cn/s/07bdc81784ce

MiniMax H3 INT8 W4A8 Workflow

01_minimax-h3-int8-w4a8-workflow.json

v1.0

<div data-youtube-video="https://youtu.be/SUJkkT6XLHI"><iframe src="https://www.youtube.com/embed/SUJkkT6XLHI" width="853" height="480" allowfullscreen></iframe></div> 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2086411704518406146?inviteCode=rh-v1111
<h2>MiniMax H3 Int8 W4A8 Acceleration Comparison — MiniMax H3 INT8 W4A8 Acceleration</h2>
A MiniMax H3 video graph from episode 291, part of the six-times-acceleration series that compares model tier, LoRA, attention and cache settings.
<h2>Saved configuration</h2>
62 nodes / 78 links. Three parallel branches saved in one file: W4A8 mixed UNET <code>minimax_h3_fl2va_pruned_w4a8_mixed.safetensors</code>, plus two <code>minimax_h3_fl2va_int8_convrot.safetensors</code> branches; CLIP <code>qwen3vl_32b_minimax_h3_nvfp4_awq</code> / <code>..._int8_convrot</code>; int8 and fp32 VAEs. No LoRA nodes. Sampling: <code>res_multistep</code> + simple, 20 steps, denoise 1, three LoadImage inputs and three outputs.
<h2>How to use</h2>
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.
<h2>Limits and not-verified items</h2>
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: <a href="https://www.youtube.com/watch?v=SUJkkT6XLHI" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV17nuU6xE3g" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.

MiniMax H3 INT8 W4A8 Acceleration

<div data-youtube-video="https://youtu.be/SUJkkT6XLHI"></div> 
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 <code>minimax_h3_fl2va_pruned_w4a8_mixed.safetensors</code>, plus two <code>minimax_h3_fl2va_int8_convrot.safetensors</code> branches; CLIP <code>qwen3vl_32b_minimax_h3_nvfp4_awq</code> / <code>..._int8_convrot</code>; int8 and fp32 VAEs. No LoRA nodes. Sampling: <code>res_multistep</code> + 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: <a href="https://www.youtube.com/watch?v=SUJkkT6XLHI" rel="ugc">YouTube</a> · <a href="https://www.bilibili.com/video/BV17nuU6xE3g" rel="ugc">Bilibili</a>. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.

Files with SHA256: 7379497ed342c3a358652600c33dac54687658cd66eac71aae69f6b7030e8e3b