MiniMax H3 (Comfy) / imageToVideo

If the plain MiniMax H3 node is the API version - 2K, aspect ratio presets, a reference stack - this one is the Comfy version of the same model: width, height, steps, seed, LoRAs, and two booleans that tell you exactly what Civitai is doing under the hood. fast "enables EasyCache with its default configuration" and turbo toggles the distilled/speed path. That's the vocabulary of local ComfyUI work, rented on a cloud GPU. For anyone who learned video gen from running LTX or Wan locally, this node will feel immediately familiar - the parameters map straight onto what you already know.

It's part of civitai-comfy-nodes, Civitai's official ~160-node pack generated from its Orchestration API spec. Cloud job: submit a videoGen, long-poll, get a native VIDEO tensor. Billed in Buzz.

Inputs that matter

prompt is the only required input. Then:

• first_frame / last_frame (IMAGE sockets) - bookend the clip. The image-to-video behavior lives in these; with neither wired you're effectively doing text-to-video at a chosen size.
• width (864) / height (480) - H3's default in Comfy form is smaller than the 2K cloud node; you trade resolution for speed and control.
• duration - default 5.
• seed - for reproduction; steps - default 20.
• fast - EasyCache on; a speed-vs-quality trade you can flip per job.
• turbo - the distilled fast path. If fast is "quicker," turbo is "quickest."
• diffusion_model (CIVITAI_AIR socket) and loras (CIVITAI_LORAS socket) - point the job at a specific H3 checkpoint and stack community LoRAs via the Civitai selectors.

Outputs: video (VIDEO), draft_cache_audio (AUDIO), draft_cache_video (VIDEO), workflow_id and raw_json (STRING).

Install and auth

Standard for the pack - official, Comfy Registry:

# ComfyUI Manager: search "Civitai Comfy Nodes"
cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Restart, authenticate via a Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth.

Common issues

fast and turbo are quality levers, not free wins. EasyCache and the distilled path both trade fidelity for latency; they're for iterating on composition, and you'll see the difference on fine detail. 864×480 is the floor, not a bug - the Comfy variant deliberately runs smaller than the 2K cloud node; don't expect cloud-node resolution here. The two H3 nodes are easy to mix up - same model, totally different parameter sets; if you're looking for referenceToVideo style multi-image conditioning, that's the sibling node, not this one. Everything leaves your machine and bills per output second, as with all cloud jobs. And the pack is early preview - behavior can change without notice.