What is this?
H3 Lite is a Codex/WorkBuddy AI agent skill that drives local MiniMax H3 video generation through ComfyUI, picking a hardware-aware route and producing videos with native audio from a description.
H3 Lite is a Codex/WorkBuddy AI agent skill that drives local MiniMax H3 video generation through ComfyUI, picking a hardware-aware route and producing videos with native audio from a description.Is this a standalone app or a skill for an existing AI agent?It is an Agent Skill — a folder of instructions, scripts and ComfyUI workflows that Codex or WorkBuddy loads to orchestrate local video generation, not a standalone application.
It is an Agent Skill — a folder of instructions, scripts and ComfyUI workflows that Codex or WorkBuddy loads to orchestrate local video generation, not a standalone application.How do I install the skill and get it running?Download the repo ZIP, unzip, and drop the h3lite folder into Codex's skills directory, then reopen Codex; ComfyUI and one of the model component sets (Set A or Set B) must be installed separately in their own folder.
Download the repo ZIP, unzip, and drop the h3lite folder into Codex's skills directory, then reopen Codex; ComfyUI and one of the model component sets (Set A or Set B) must be installed separately in their own folder.Do I need a specific GPU or OS to run this?Yes — the only fully verified route is Windows with an NVIDIA GPU running ComfyUI; the skill depends on either the Set A (W4A8 + 4B INT4 encoder) or Set B (4B FP8 encoder) component package.
Yes — the only fully verified route is Windows with an NVIDIA GPU running ComfyUI; the skill depends on either the Set A (W4A8 + 4B INT4 encoder) or Set B (4B FP8 encoder) component package.How well does it run on lower-end hardware?Speed varies a lot even at the same settings: the identical Set B workflow at 640x352/4 steps takes about 77 seconds on an RTX 4060 Ti 16GB versus about 591 seconds on an RTX 4070 Laptop 8GB.
Speed varies a lot even at the same settings: the identical Set B workflow at 640x352/4 steps takes about 77 seconds on an RTX 4060 Ti 16GB versus about 591 seconds on an RTX 4070 Laptop 8GB.Can I use this on Mac or Linux?Only unofficially — macOS Apple Silicon is a community route that points to a separate MLX/Metal port, macOS Intel is unverified, and Linux + NVIDIA is an experimental route requiring manual path and node adaptation.
Only unofficially — macOS Apple Silicon is a community route that points to a separate MLX/Metal port, macOS Intel is unverified, and Linux + NVIDIA is an experimental route requiring manual path and node adaptation.What resolutions and video lengths does it support?Default fast mode renders 640x352 clips with native audio in 4 steps; documented presets range from a 608x352 low-VRAM preview up to roughly 1920x1056 (1080p), with higher resolutions recommended only after a low-res preview run.
Default fast mode renders 640x352 clips with native audio in 4 steps; documented presets range from a 608x352 low-VRAM preview up to roughly 1920x1056 (1080p), with higher resolutions recommended only after a low-res preview run.Can it upscale videos to 1080p or 4K afterward?Yes, as a separate post-processing step — finished clips can be handed off to a locally installed Topaz Video AI (Starlight/Astra Fast, Proteus or Rhea) or a FlashVSR CLI setup, but the skill never substitutes upscaling for regenerating the video itself.
Yes, as a separate post-processing step — finished clips can be handed off to a locally installed Topaz Video AI (Starlight/Astra Fast, Proteus or Rhea) or a FlashVSR CLI setup, but the skill never substitutes upscaling for regenerating the video itself.Can I track generation progress without opening a web browser?On Windows, the fastpath launcher pops up a native progress window that reads ComfyUI's own WebSocket feed to show queueing, sampling and decoding steps plus elapsed time, ETA, VRAM, RAM and pagefile use.
On Windows, the fastpath launcher pops up a native progress window that reads ComfyUI's own WebSocket feed to show queueing, sampling and decoding steps plus elapsed time, ETA, VRAM, RAM and pagefile use.What kinds of inputs can it turn into video — just text, or images and reference clips too?Four workflows cover different inputs: pure text (T2VA), a locked first frame (I2VA), matched first/last frames (FL2VA), and multi-reference images, video or audio (Ref2VA), each needing differently structured prompts.
Four workflows cover different inputs: pure text (T2VA), a locked first frame (I2VA), matched first/last frames (FL2VA), and multi-reference images, video or audio (Ref2VA), each needing differently structured prompts.What are similar projects?FreeVideo, CodeVideoModel, ComfyUI-ALLinONE-MinimaxH3, God-minmax-H3, seedance-2.0, MiniMax-H3, short-drama-production, MiniMax-H3-Cloud
FreeVideo, CodeVideoModel, ComfyUI-ALLinONE-MinimaxH3, God-minmax-H3, seedance-2.0, MiniMax-H3, short-drama-production, MiniMax-H3-CloudWhich topics is it in?ai-video-generation, minimax-h3, agentic-workflow, upscaling, image-to-video, comfyui, text-to-video, ffmpeg, video-generation, prompt-engineering, skill, nvidia
ai-video-generation, minimax-h3, agentic-workflow, upscaling, image-to-video, comfyui, text-to-video, ffmpeg, video-generation, prompt-engineering, skill, nvidia