H3 Lite: a Codex skill that picks a ComfyUI route for local MiniMax H3 video

The gap H3 Lite fills between a prompt and a ComfyUI graph

MiniMax H3 is a video model that runs locally through ComfyUI, and getting it to produce anything means assembling a diffusion model, a text encoder, a ClipProj, two VAEs, a Turbo LoRA and a set of custom nodes, then wiring them into a graph whose parameters depend on how much VRAM you have. H3 Lite is a Codex skill that absorbs that assembly work. You describe the shot; the agent reads your hardware, picks a route, prepares the components and checks the result.

The audience is narrow and the README is honest about it. The primary supported path is Windows with an NVIDIA GPU. The verified machines are a laptop with an RTX 4070 Laptop 8 GB and 32 GB of system memory, and a desktop with an RTX 4060 Ti 16 GB. If you are on macOS Apple Silicon the README points at a separate community route, `mmh3turbo` with `uvx`, and says to use it independently of this skill's ComfyUI components. macOS Intel is listed as unverified. Linux with NVIDIA is marked experimental, with the note that you adapt paths, nodes and run parameters yourself.

Intent routing, anchor cards and the two component sets

The workflow is four stages: intent routing, reference-image anchoring, prompt enhancement, then generation and acceptance. Routing is the interesting part. A request maps to one of four modes. T2VA is text only. I2VA pins a reference image at `0.00s` and describes only what changes after. FL2VA takes a first and last frame and asks for a continuous visible change between them. Ref2VA handles multiple images, video or audio references, and the README's instruction is to define each asset's role, what to preserve and what may vary before writing the shot list.

For anything with people or multiple shots, the agent builds an anchor card fixing character, wardrobe, props, scene and lighting, and gives reference assets stable names like `Subject A` and `Picture 1`. At runtime the card is written to `anchors.json` and its path recorded in `manifest.json`. An `anchor_qa` step checks first, middle and last frames for continuity. The README is explicit that identity, clothing and composition still need a human pass. If the Ref2VA components are incomplete, the agent falls back to I2VA or flags the request as experimental.

Prompts are expanded in a fixed order (intent, scene and character, action and shot list, camera and sound, anti-drift constraints) and then converted into H3's `integrated_multimodal_description`, `overall_soundscape` and `non_diegetic_music` fields. The component side splits into Set A, built on a W4A8 diffusion model with a 4B INT4 text encoder for the low-VRAM fast route, and Set B, which uses a 4B FP8 text encoder for FP8 compatibility. You pick one; they are not meant to be mixed.

Installing H3 Lite and running the red-ball check

The README's install path is conversational rather than scripted. You send the repository URL to Codex or WorkBuddy and ask it to prepare the environment for your machine. If you want the models off your system drive, you put the target directory in the same message, as in the second example below.

请帮我安装 H3 Lite，并根据我的电脑配置准备本地 MiniMax H3 视频生成环境：
https://github.com/Rimagination/h3lite

请把 MiniMax H3 和 ComfyUI 安装到 F:\MiniMax-H3；如果那里已经有健康环境就直接复用。

Before downloading, decide where ComfyUI lives. The README offers three placements: reuse an existing `<ComfyUI>` tree to keep your current models and nodes, use a standalone directory such as `F:\MiniMax-H3\ComfyUI` to avoid filling the system disk, or put it under `<项目>\.h3lite\ComfyUI` so the environment travels with the project. Then download one component set, merge its `models` and `custom_nodes` into `<ComfyUI>`, import the workflow JSON and keep `component-manifest.json` alongside it. Set A and Set B are distributed through Baidu share links; the README does not give an alternative mirror for either.

If you would rather not use the agent for installation, the manual route is Code → Download ZIP on the repository page, then place the extracted `h3lite` folder in Codex's skills folder and reopen Codex.

The first real test is deliberately small: a 5 second landscape clip with simple motion and unambiguous sound, so the whole chain (model, nodes, audio, output directory) is exercised in one pass. The README supplies the prompt, a matte red rubber ball bouncing twice on grey concrete and rolling out of frame, shot on a low fixed camera in cold overcast light, keeping the two impacts and the rolling sound with no music. Check the video, the motion and the native audio before raising the canvas or the complexity.

What the 8 GB route actually costs you

The README publishes one comparison that matters more than any feature list. On the same Set B, compatible workflow, prompt, seed and `640×352 / 4 步` parameters, the RTX 4060 Ti 16 GB finished in about 77.08 seconds and the RTX 4070 Laptop 8 GB took about 591.22 seconds. That is roughly a sevenfold gap for identical settings, and it is the single most useful number in the repository because it tells you what the low-VRAM route buys and what it charges.

That shapes the tool's fit. On 8 GB, the fast preset (4 steps, native audio, 640×352) is the realistic working mode, and the README's own guidance is to preview at low resolution first and raise the canvas afterwards. The resolution table goes from 608×352 at 0.21 MP up to 1920×1056 at 2.03 MP, with 864×480 as the common 16:9 quality step. Nothing in the README claims those higher tiers are comfortable on 8 GB. VRAM, system memory, pagefile, disk and laptop power limits all affect speed, and only two machines are listed as verified.

A second limitation is the acceptance step. `anchor_qa` checks frames, but the README states plainly that identity, wardrobe and composition still require human review. If your task depends on a specific face staying consistent across shots, the tool narrows the search space rather than removing the check. Treat it as a first-pass generator, not an unattended pipeline.

H3 Lite against the mmh3turbo Mac route

The genuine alternative named in the README is the community Mac path: the `mmh3turbo` bundles on Hugging Face, run through `uvx mmh3turbo`. The difference is architectural, not cosmetic. H3 Lite drives ComfyUI graphs and ships matched component sets (models, nodes, workflows, manifest) tuned for NVIDIA VRAM tiers, with a planner and a doctor in the repo. The Mac route targets MLX and Metal instead, and the README instructs you to use it separately from this skill's ComfyUI components, which means no shared workflow JSON and no shared component manifest.

So the choice is not which is faster; the README publishes no cross-platform timing. It is which backend your hardware can run at all. If you are on Apple Silicon, H3 Lite's primary path does not apply to you and the community route is the one the README points to. If you are on Windows with an NVIDIA card, the ComfyUI route is the supported one and the Mac bundles are irrelevant. Linux users sit outside both: experimental here, with manual adaptation expected.

Within the Windows path there is a second fork worth knowing. Set A is the low-VRAM fast route and Set B is the FP8 compatibility route. The README's Set A gallery compares six acceleration-node combinations (compatibility baseline, Sage only, FFN only, Block Cache only, Sol only, and all accelerators together) on one identical rolling-red-ball scene at `640×352 / 4 步` with native audio, and the six MP4s are downloadable from the `assets-rolling-redball` release. That is a route-comparison artifact, not a quality ranking.

Maintenance, licence and the upgrade question

The repository is not archived, and the last push was on 2026-08-23. The two most recent releases, `assets` and `assets-rolling-redball`, both landed on 2026-08-20 and both are showcase material rather than installers: the second is described as a download backup for the six Set A route videos. So the release history tells you about gallery assets, not about component updates.

The upgrade story is the weak point. The README documents how to install the skill and how to merge a component set into ComfyUI, but it does not document how to move from one component set version to the next, how to roll back a workflow JSON that a new node pack breaks, or how to reconcile a changed `component-manifest.json` with an existing `anchors.json` and `manifest.json` in your output tree. Anyone running this in a shared environment should assume they will re-derive that procedure themselves.

The licence is MIT, which is permissive and places few obligations on how you redistribute or modify the skill's own code. Two caveats the licence text does not settle and this article will not either: the component sets are distributed through Baidu share links rather than the repository, so the terms attached to the model weights and custom nodes come from their upstream projects, not from H3 Lite's MIT file. Check those upstream licences separately before any commercial use. For legal advice, ask a lawyer.

Editorial conclusion

Adopt H3 Lite if you already run ComfyUI on Windows with an NVIDIA card and want an agent to pick the route, stage the components and hand back an MP4 with native audio. Skip it if you are on macOS Intel, on a Linux box you are not prepared to adapt by hand, or if you need a maintained release cadence you can plan around: the last push was on 2026-08-23 and the README does not document an upgrade path. Verify three things before you commit a disk: which component set your VRAM actually tolerates, where the ComfyUI tree will live, and whether the Baidu share links for Set A and Set B are still reachable.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. Rimagination/h3lite on GitHub